def image_proposal(img_path):
img = skimage.io.imread(img_path)
img_lbl, regions = selectivesearch.selective_search(
img, scale=500, sigma=0.9, min_size=10)
candidates = set()
images = []
vertices = []
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
if r['size'] < 220:
continue
# resize to 224 * 224 for input
proposal_img, proposal_vertice = prep.clip_pic(img, r['rect'])
# Delete Empty array
if len(proposal_img) == 0:
continue
# Ignore things contain 0 or not C contiguous array
x, y, w, h = r['rect']
if w == 0 or h == 0:
continue
# Check if any 0-dimension exist
[a, b, c] = np.shape(proposal_img)
if a == 0 or b == 0 or c == 0:
continue
im = Image.fromarray(proposal_img)
resized_proposal_img = resize_image(im, 224, 224)
candidates.add(r['rect'])
img_float = pil_to_nparray(resized_proposal_img)
images.append(img_float)
vertices.append(r['rect'])
return images, vertices
def image_proposal(img_path):
'''
Enter the picture to be candidate box extraction
The candidate box is extracted by using the characteristics of each pixel of the picture,
because the number of candidate boxes is too many and the size of candidate boxes required for different problem backgrounds is not the same.
Therefore, it is necessary to go through a series of rules to further reduce the number of feature frames.
'''
img = cv2.imread(img_path)
img_lbl, regions = selectivesearch.selective_search(img, scale=500, sigma=0.9, min_size=10)
candidates = set()
images = []
vertices = []
for r in regions:
if r['rect'] in candidates:
continue
if r['size'] < 220:
continue
if (r['rect'][2] * r['rect'][3]) < 500:
continue
proposal_img, proposal_vertice = clip_pic(img, r['rect'])
if len(proposal_img) == 0:
continue
x, y, w, h = r['rect']
if w == 0 or h == 0:
continue
[a, b, c] = np.shape(proposal_img)
if a == 0 or b == 0 or c == 0:
continue
resized_proposal_img = resize_image(proposal_img, cfg.Image_size, cfg.Image_size)
candidates.add(r['rect'])
img_float = np.asarray(resized_proposal_img, dtype="float32")
images.append(img_float)
vertices.append(r['rect'])
return images, vertices
def main():
# loading astronaut image
# img = skimage.data.coffee()
img = np.asarray(Image.open('berlin2.jpg'))
# perform selective search
img_lbl, regions = selectivesearch.selective_search(img,
scale=700,
sigma=0.8,
min_size=10)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 800:
continue
# distorted rects
x, y, w, h = r['rect']
# if w / h > 1.2 or h / w > 1.2:
# continue
candidates.add(r['rect'])
# draw rectangles on the original image
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
f = open(time.ctime(), 'wr')
pickle.dump(candidates, f)
f.close()
def predict(img, pngFilePath=None):
# loading lena image
# img = skimage.data.chelsea() # => numpy.ndarray
# print img[0][0]
# 書きの方法はダメ！ネガポジ画像になってしまう。
# img = Image.open("images/fruit.jpg")
# img.load()
# img = np.asarray( img, dtype="int32" )
# predictでは、data/bbox_image/n07747607_631_box2.JPEGは正解できた。
# ということは、n07747607_631.jpg でselective search すると正解できるかも？
# tgt_img_path = "data/image/n07747607_631.jpg"
if pngFilePath is not None:
im = Image.open(pngFilePath)
bg = Image.new("RGB", im.size, (255,255,255))
bg.paste(im,im)
bg.save(pngFilePath + ".jpg")
img = io.imread(pngFilePath + ".jpg")
print "get img from file"
# perform selective search
img_lbl, regions = selectivesearch.selective_search(img, scale=500, sigma=0.9, min_size=10)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 2000:
continue
# distorted rects
x, y, w, h = r['rect']
if w / h > 1.2 or h / w > 1.2:
continue
candidates.add(r['rect'])
print candidates
# img = Image.open(tgt_img_path)
# imgs = [resize(img.crop(rect)) for rect in candidates]
res = []
img = Image.open(pngFilePath + ".jpg")
for rect in candidates:
print rect
cropped_img = resize(img.crop(rect))
if cropped_img is None:
continue
print "cropped!"
(answer, prob) = pd.predict_by_data(cropped_img)
print "get answer and prob!"
if prob > 0.9:
res.append((rect,answer))
print "send back answers!"
return res
def img_proposal(img_path):
""" 将给定的图片转换成 （框图片， 所在位置）的列表 """
# img = skimage.io.imread(img_path)
img = RCNN.load_img(img_path)
img_lbl, regions = selectivesearch.selective_search(img,
scale=500,
sigma=0.9,
min_size=10)
candidates = set()
images = list()
vertices = list()
for r in regions:
if r['rect'][2] == 0 or r['rect'][3] == 0:
continue
if r['size'] < 220:
continue
if r['rect'] in candidates:
continue
candidates.add(r['rect'])
proposal_img, proposal_vertice = clip_img(img, r['rect'])
resized_proposal_img = RCNN.resize_img(proposal_img, (224, 224))
images.append(resized_proposal_img)
vertices.append(proposal_vertice)
return images, vertices
def predict_img(img, model, place_list, input_shape):
# 画像から特徴部分を抽出
img_lbl, regions = selectivesearch.selective_search(img,
scale=500,
sigma=0.9,
min_size=10)
# 分類器を使用して判別
candidates = set()
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(img)
for r in regions:
# 特徴量抽出
if r['rect'] in candidates:
continue
# ある一定pixelより小さいものについては除外する
if r['size'] < 300:
continue
# ある一定pixelより大きいものについては除外する
if r['size'] > 750:
continue
# 特徴量部分の座標を取得
x, y, w, h = r['rect']
if 2 * h > w:
continue
if h < 5:
continue
#print("hの高さ%s" % h)
img_part = img[y:y + h, x:x + w]
ax, place_name_list = local_classification(ax, img_part, input_shape,
x, y, w, h)
ax.imshow(img)
return place_name_list
def predict_selective_search(img, n_rect=7, n_rect_search=200):
rects = selectivesearch.selective_search(img,
scale=500,
sigma=0.9,
min_size=10)[1][:n_rect_search]
centers = []
candidates = set()
for r in rects:
if r['rect'] in candidates:
continue
if r['size'] < 1200:
continue
x, y, w, h = r['rect']
candidates.add(r['rect'])
candidates = list(candidates)[:n_rect]
for box in candidates:
x, y, w, h = box
centers += [[x, y, x + w, y + h]]
if len(centers) == 0: # if no object detected, center crop
tx1 = (img.shape[1] - cfg.crop_size) / 2
ty1 = (img.shape[0] - cfg.crop_size) / 2
tx2 = (img.shape[1] + cfg.crop_size) / 2
ty2 = (img.shape[0] + cfg.crop_size) / 2
centers = np.array([[tx1, ty1, tx2, ty2]], dtype=np.float32)
return np.asarray(centers, dtype=np.float32), np.zeros(len(centers),
dtype=np.float32)
def get_regions(img_path):
# loading image
img = io.imread(img_path)
# perform selective search
img_lbl, regions = selectivesearch.selective_search(img,
scale=500,
sigma=0.9,
min_size=10)
print(len(img_lbl))
print(np.shape(img))
candidates = set()
for r in regions:
# excluding same rectangle (withsl different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 100:
continue
candidates.add(r['rect'])
# draw rectangles on the original image
for i, R in enumerate(candidates):
x, y, w, h = R
nor_img = transform.resize(img[y:y + h, x:x + w, :], (224, 224))
io.imsave("./region/" + str(i) + '.jpg', nor_img)
def visualize(img):
# loading astronaut image
# img = skimage.data.astronaut()
# perform selective search
img_lbl, regions = selectivesearch.selective_search(
img, scale=500, sigma=0.9, min_size=10)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 220:
continue
# distorted rects
x, y, w, h = r['rect']
# if w / h > 1.2 or h / w > 1.2:
# continue
candidates.add(r['rect'])
# draw rectangles on the original image
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(img)
for i, r in enumerate(candidates):
x, y, w, h = r
print(i, x, y, w, h)
rect = mpatches.Rectangle(
(x, y), w, h, fill=False, edgecolor='red', linewidth=1)
ax.add_patch(rect)
plt.show()
def main():
# loading astronaut image
#img = skimage.data.astronaut()
# loading your own image
img = io.imread('./sheep.jpg')
# perform selective search
img_lbl, regions = selectivesearch.selective_search(img,
scale=100,
sigma=0.8,
min_size=10)
print(regions[0])
"""
parameter explanation
sigma:
In general we use a Gaussian filter to smooth the image slightly before
computing the edge weights, in order to compensate for digitization artifacts.
We always use a Gaussian with σ = 0.8, which does not produce any
visible change to the image but helps remove artifacts.
(The smaller the sigma, the finer the segmentation result)
min_size:
If the rect size is reached on min_size, the calculation is stopped.
scale:
a larger scale causes a preference for larger components
"""
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions lager than 1600 pixels
if r['size'] > 1600 or r['size'] < 50:
continue
# distorted rects
x, y, w, h = r['rect']
if w / h > 1.2 or h / w > 1.2:
continue
candidates.add(r['rect'])
# draw rectangles on the original image
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(img)
for x, y, w, h in candidates:
#print(x, y, w, h)
rect = mpatches.Rectangle((x, y),
w,
h,
fill=False,
edgecolor='red',
linewidth=1)
ax.add_patch(rect)
plt.show()
def selectiveSearch(img, pos, template):
"""
Args:
Image: the orginal image. Used for selective search
Pos: (x,y,w,h) of the bounding box
Template: the template captured in the first frame
Output:
proposed regions
"""
x, y, w, h = pos
min_size = np.floor(w * h / 4).astype(int)
img_lbl, regions = selectivesearch.selective_search(img,
scale=500,
sigma=0.9,
min_size=min_size)
candidates = []
rect_img = img.copy()
for item in regions:
rect = item['rect']
pt1 = rect[0], rect[1]
pt2 = rect[0] + rect[2], rect[1] + rect[3]
cv2.rectangle(rect_img, pt1, pt2, (0, 255, 0), 2)
if rect[2] > 1.75 * h or rect[3] > 1.75 * w or rect[
2] < 0.25 * h or rect[3] < 0.25 * w:
continue
if np.abs(rect[0] -
y) > img.shape[1] / 1.5 or np.abs(rect[1] -
x) > img.shape[0] / 1.5:
continue
candidates.append(rect)
#cv2.imshow("test", rect_img)
#cv2.waitKey(0)
best_rect = findPatch(img, candidates, template)
return best_rect
def TestRCNN(self, input_image):
predicted_classes = []
predicted_bboxes = []
if type(input_image) == type(''):
input_image = skimage.io.imread(input_image)
_, sc_regions = selectivesearch.selective_search(input_image)
proposal_regions = []
for region in sc_regions:
proposal_regions.append([
region['rect'][0], region['rect'][1],
region['rect'][2] - region['rect'][0],
region['rect'][3] - region['rect'][1]
])
warped_region = []
for region in proposal_regions:
warped_region.append(self.RegionWarp(input_image, region))
for region in warped_region:
bbox, class_ = self.TOTAL_model.predict(region)
predicted_bboxes.append(bbox)
predicted_classes.append(class_)
predict_results = []
for idx in range(len(predicted_bboxes)):
predict_results.append([
predicted_bboxes[idx][0], predicted_bboxes[idx][1],
predicted_bboxes[idx][2], predicted_bboxes[idx][3],
predicted_classes[idx]
])
return predict_results
def get_object_proposals(img, scale=500, sigma=0.9, min_size=10):
# Selective search
img_lbl, regions = selectivesearch.selective_search(img,
scale=scale,
sigma=sigma,
min_size=min_size)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 500 pixels
x, y, w, h = r['rect']
if r['size'] < 500 or w > 0.95 * img.shape[1] or h > 0.95 * img.shape[
0]:
continue
# excluding the zero-width or zero-height box
if r['rect'][2] == 0 or r['rect'][3] == 0:
continue
# distorted rects
if w / h > 5 or h / w > 5:
continue
candidates.add(r['rect'])
return candidates
def customized_selective_search(img):
assert img.shape[-1] == 3, "imgs.shape should be [B, H, W, 3]"
ss_boxes = []
img_lbl, regions = selectivesearch.selective_search(img,
scale=500,
sigma=0.9,
min_size=10)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 500 pixels
if r['size'] < 50:
continue
x, y, w, h = r['rect']
if w < 5 or h < 5 or w / h > 2.0 or h / w > 2.0:
continue
candidates.add(r['rect'])
data = [list(a) for a in candidates]
data = np.array(data)
if data.shape[0] == 0:
print("this image has no proposals")
else:
data[:, 2:] = data[:, :2] + data[:, 2:]
ss_boxes.append(list(data))
return ss_boxes
def load_regions(image_index, number):
image_name = image_index + ".jpg"
xml_name = image_index + ".xml"
image_dir = os.path.join(IMAGE_DIR, image_name)
xml_dir = os.path.join(ANNOTATION_DIR, xml_name)
# get regions proposals with selective search
image = skimage.io.imread(image_dir)
_, regions = selectivesearch.selective_search(image)
# get ground truth
dom_tree = xml.dom.minidom.parse(xml_dir)
annotation = dom_tree.documentElement
object_list = annotation.getElementsByTagName('object')
for object in object_list:
dom_name = object.getElementsByTagName('name')
object_name = dom_name[0].childNodes[0].data
xmin = int(object.getElementsByTagName('xmin')[0].childNodes[0].data)
ymin = int(object.getElementsByTagName('ymin')[0].childNodes[0].data)
xmax = int(object.getElementsByTagName('xmax')[0].childNodes[0].data)
ymax = int(object.getElementsByTagName('ymax')[0].childNodes[0].data)
ground_truth = [xmin, ymin, xmax, ymax]
region_proposal = []
background = []
for region in regions:
roi = region['rect']
iou = utils.get_IoU(ground_truth, roi)
if iou > 0.5 and len(region_proposal) < number * 0.25:
label = int(classes_num[classes == object_name])
region_proposal.append([roi, [label, ground_truth]])
elif iou > 0.1 and len(background) < number * 0.75:
background.append([roi, [20, -1]])
if len(region_proposal) is 16 and len(background) is 48:
break
rois = region_proposal + background
return image, rois
def pp_nd_ss(image_dir):
ss_arr = []
img = skimage.io.imread(image_dir)
img = Image.fromarray(img).resize((640, 480))
img = np.array(img)
img_lbl, regions = selectivesearch.selective_search(img,
scale=500,
sigma=0,
min_size=500)
candidates = []
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 2000:
continue
# distorted rects
x, y, w, h = r['rect']
if h is 0 or w is 0:
continue
if w / h > 2 or h / w > 2:
continue
candidates.append(r['rect'])
image = Image.fromarray(img).crop((x, y, x + w, h + y)).resize(
(64, 64))
ss_arr.append(np.array(image))
ss_arr = np.array(ss_arr) / 255
return ss_arr, candidates
def generate_labels(self):
with codecs.open(self.all_list, 'r', 'utf-8') as f:
lines = f.readlines()
for num, image_idx in enumerate(lines):
ground_truth_dic = self.load_annotation(image_idx)
image_path = os.path.join(self.images_path,
image_idx.strip() + '.jpg')
img = cv2.imread(image_path)
img_lbl, regions = selectivesearch.selective_search(
img, scale=1000, sigma=0.9, min_size=1000)
labels = []
for r in regions:
x, y, w, h = r['rect']
proposal_vertice = [x + 1, y, x + w, y + h, w, h]
proposal_bbox = [x, y, (x + w - 1), (y + h - 1)]
label = np.zeros(self.class_num * 5 - 4, dtype=np.float32
) # 假设包括背景有5类，0：5是判断类别，5：5+4*4=21 是位置框信息
iou_val = 0
for ground_truth, class_idx in ground_truth_dic.items():
#ground_truth = list(ground_truth)
xmin = (2 * ground_truth[0] - ground_truth[2]) / 2.0
ymin = (2 * ground_truth[1] - ground_truth[3]) / 2.0
ground_truth = [
xmin, ymin, ground_truth[2], ground_truth[3]
]
iou_val = IOU(ground_truth, proposal_bbox)
px = float(proposal_vertice[0]) + float(
proposal_vertice[4] / 2.0) # 中心点X
py = float(proposal_vertice[1]) + float(
proposal_vertice[5] / 2.0) # 中心点Y
pw = float(proposal_vertice[4]) # w
ph = float(proposal_vertice[5]) # h
gx = float(ground_truth[0]) # 中心点X
gy = float(ground_truth[1]) # 中心点Y
gw = float(ground_truth[2]) # W
gh = float(ground_truth[3]) # H
if iou_val < self.roi_threshold:
label[0] = 1
elif iou_val > self.roi_threshold:
label[0] = 0
label[class_idx] = 1
label[self.class_num + (class_idx-1)*4 : self.class_num + (class_idx-1)*4 + 4] = \
[((gx - px) / pw), ((gy - py) / ph), (np.log(gw / pw)), (np.log(gh / ph))]
break
for i in range(len(proposal_bbox)):
proposal_bbox[i] = (proposal_bbox[i] / 16.0)
proposal_bbox.insert(0, 0)
proposal_bbox.insert(0, iou_val)
proposal_bbox.extend(label)
labels.append(proposal_bbox)
view_bar("Process image of %s" % image_path, num + 1,
len(lines))
if self.is_save:
if not os.path.exists(self.processed_path):
os.makedirs(self.processed_path)
np.save((os.path.join(self.processed_path,
image_idx.split('.')[0].strip()) +
'_data.npy'), labels)
def extract_regions(image, scale=500, min_pixels=1000):
'''
img: np.narray in tf format
return: list [[x, y, w, h], ...,]
'''
img = image.copy()
img = img.astype('uint8') # for fast
img = np.rollaxis(img, 0, 3) # change to dim_ordering to tf
# get roi
img_lbl, region = sls.selective_search(img,
scale=scale,
sigma=0.8,
min_size=50)
candidates = set()
for r in region:
# excluding same rectangle (with different segments)
if r['rect'] in set():
continue
# excluding regions smaller than 2000 pixels
if r['size'] < min_pixels:
continue
candidates.add(r['rect'])
return list(candidates)
def _get_ssinfo_(image):
# 单通道 to 三通道
image = np.expand_dims(image, axis=2)
image = np.concatenate((image, image, image), axis=-1)
img_lbl, regions = selectivesearch.selective_search(image,
scale=100,
sigma=0.1,
min_size=10)
# 计算一共分割了多少个原始候选区域
temp = set()
for i in range(img_lbl.shape[0]):
for j in range(img_lbl.shape[1]):
temp.add(img_lbl[i, j, 3])
# 创建一个集合 元素不会重复，每一个元素都是一个list(左上角x，左上角y,宽,高)，表示一个候选区域的边框
candidates = set() # 注意，set只能遍历
for r in regions:
x, y, w, h = r['rect']
# 排除小于 2000 pixels的候选区域(并不是bounding box中的区域大小)
if r['size'] > 120:
continue
# 排除扭曲的候选区域边框 即只保留近似正方形的
x, y, w, h = r['rect']
if w == 0 or h == 0:
continue
try:
if w / h < 0.25 or h / w < 0.25:
continue
except ZeroDivisionError:
print(1)
continue
candidates.add(r['rect'])
return candidates
def main():
# loading lena image
#img = skimage.data.lena()
img = skimage.data.imread("moto.png")
# perform selective search 500 0.9 10
img_lbl, regions = selectivesearch.selective_search(
img, scale=400, sigma=0.9, min_size=10)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 2000:
continue
# distorted rects
x, y, w, h = r['rect']
if w / h > 1.2 or h / w > 1.2:
continue
candidates.add(r['rect'])
# draw rectangles on the original image
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(img)
for x, y, w, h in candidates:
print x, y, w, h
rect = mpatches.Rectangle(
(x, y), w, h, fill=False, edgecolor='red', linewidth=1)
ax.add_patch(rect)
plt.show()
def image_proposal(img_path):
'''
输入要进行候选框提取的图片
利用图片的各像素点的特点进行候选框的提取，由于候选框数量太多且针对不同的问题背景所需要的候选框的尺寸是不一样的
因此要经过一系列的规则加以限制来进一步减小特征框的数量
'''
img = cv2.imread(img_path)
img_lbl, regions = selectivesearch.selective_search(img, scale=500, sigma=0.9, min_size=10)
candidates = set()
images = []
vertices = []
for r in regions:
if r['rect'] in candidates:
continue
if r['size'] < 220:
continue
if (r['rect'][2] * r['rect'][3]) < 500:
continue
proposal_img, proposal_vertice = clip_pic(img, r['rect'])
if len(proposal_img) == 0:
continue
x, y, w, h = r['rect']
if w == 0 or h == 0:
continue
[a, b, c] = np.shape(proposal_img)
if a == 0 or b == 0 or c == 0:
continue
resized_proposal_img = resize_image(proposal_img, cfg.Image_size, cfg.Image_size)
candidates.add(r['rect'])
img_float = np.asarray(resized_proposal_img, dtype="float32")
images.append(img_float)
vertices.append(r['rect'])
return images, vertices
def _generate_region_proposals(img):
"""generate region proposals"""
_, regions = selectivesearch.selective_search(img,
scale=500,
sigma=0.9,
min_size=100)
return regions
def preprocess_data(file_path, num_classes, save_path="./select_img_data",
svm=False, save=False, img_size=224, threshold=0.5):
with open(file_path, "r") as f:
lines = f.readlines()
i = 0
for line in lines:
labels = []
img_datas = []
temp = line.strip().split(" ")
img_path = temp[0]
img_label = int(temp[1])
img_rect = [int(i) for i in temp[2].split(",")]
img_data = cv2.imread(img_path)
# img_data_np = np.array(img_data, dtype="float32")
select_img_datas, select_img_rects = selectivesearch.selective_search(
img_data, scale=500, sigma=0.9, min_size=10)
rects = set()
for r in select_img_rects:
if r["rect"] in rects:
continue
if r['size'] < 220:
continue
if (r['rect'][2] * r['rect'][3]) < 500:
continue
x, y, w, h = r["rect"]
if w == 0 or h == 0:
continue
select_img_data = img_cutting(img_data, r["rect"])
if len(select_img_data) == 0:
continue
[a, b, c] = np.shape(select_img_data)
if a == 0 or b == 0 or c == 0:
continue
rects.add(r["rect"])
try:
select_img_data_resize = cv2.resize(select_img_data, (img_size, img_size), cv2.INTER_CUBIC)
except:
print(select_img_data.shape)
select_img_data_resize_np = np.array(select_img_data_resize, dtype="float32")
img_datas.append(select_img_data_resize_np)
label = np.zeros(num_classes)
iou_val = iou(r["rect"], img_rect)
if svm:
if iou_val < threshold:
labels.append(0)
else:
labels.append(img_label)
else:
if iou_val < threshold:
label[0] = 1
else:
label[img_label] = 1
labels.append(label)
print("\rreading file: %d" % i, end='', flush=True)
i += 1
if save:
if not os.path.exists(save_path):
os.mkdir(save_path)
with open((os.path.join(save_path, img_path.split('/')[-1].split('.')[0].strip()) + '.pkl'), "wb") as f:
pkl.dump((img_datas, labels), f)
def regionProposal(image_filename, size, scale=250, sigma=1):
true_image = skimage.io.imread(image_filename)
if true_image is None:
print("CANNOT FIND IMAGE TO CONDUCT DETECTION")
return None
img = cv2.resize(
true_image,
(300,
int(300.0 * float(true_image.shape[0]) / float(true_image.shape[1]))),
interpolation=cv2.INTER_AREA)
img_size = true_image.shape
boundingBoxInfo = []
images = []
aspect_ratio = float(img_size[1]) / 300.0
img_lbl, regions = selectivesearch.selective_search(img,
scale=scale,
sigma=sigma)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# # # excluding regions smaller than 400 pixels
if r['size'] < 400:
continue
# # distorted rects
x, y, w, h = r['rect']
if (w > 280 or h > 280) or (w < 15 or h < 15):
continue
candidates.add(r['rect'])
for x, y, w, h in candidates:
window = img[y:y + h, x:x + w]
window = cv2.resize(window, (size, size), interpolation=cv2.INTER_AREA)
images.append(window)
new_x = int(x * aspect_ratio)
new_y = int(y * aspect_ratio)
new_w = int(w * aspect_ratio)
new_h = int(h * aspect_ratio)
boundingBoxInfo.append([new_x, new_y, new_w, new_h])
cv2.rectangle(true_image, (new_x, new_y),
(new_x + new_w, new_y + new_h), (0, 0, 255), 1)
# cv2.imshow("Window", true_image)
# cv2.waitKey(1000)
# time.sleep(10)
return [true_image, np.array(images), boundingBoxInfo]
def get_component_position(img, scale, sigma, min_size, is_using_thumb=True):
thumb_img = img
scale_index = 1
if is_using_thumb:
pixels = 300 * 400
original_w = img.shape[0]
original_h = img.shape[1]
original_pixels = original_w * original_h
scale_index = math.sqrt(original_pixels / pixels)
new_w = int(original_w / scale_index)
new_h = int(original_h / scale_index)
thumb_img = transform.resize(img, (new_w, new_h))
# perform selective search
img_lbl, regions = selectivesearch.selective_search(
thumb_img, scale=scale, sigma=sigma, min_size=min_size)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 200:
continue
# distorted rects
x, y, w, h = r['rect']
if h == 0 or w == 0:
continue
# if w / h > 4 or h / w > 4:
# continue
candidates.add(r['rect'])
if is_using_thumb:
candidates = {tuple(int(loc * scale_index) for loc in candidate) for candidate in candidates}
# print(candidates)
return candidates
def main():
img = io.imread('cat.jpg')
# perform selective search
img_lbl, regions = selectivesearch.selective_search(
img, scale=500, sigma=0.9, min_size=10)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 1000:
continue
# distorted rects
x, y, w, h = r['rect']
if w / h > 1.2 or h / w > 1.2:
continue
candidates.add(r['rect'])
# draw rectangles on the original image
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(img)
for x, y, w, h in candidates:
rect = mpatches.Rectangle(
(x, y), w, h, fill=False, edgecolor='red', linewidth=1)
ax.add_patch(rect)
plt.show()
def selective_search(img_path):
# https://github.com/AlpacaDB/selectivesearch
img = cv2.imread(img_path)
img_lbl, regions = selectivesearch.selective_search(img,
scale=500,
sigma=0.9,
min_size=10)
# 创建一个集合 元素不会重复，每一个元素都是一个list(左上角x，左上角y,宽,高)，表示一个候选区域的边框
candidates = set()
for r in regions:
# 排除重复的候选区
if r['rect'] in candidates:
continue
# 排除小于 2000 pixels的候选区域(并不是bounding box中的区域大小)
if r['size'] < 2000:
continue
# 排除扭曲的候选区域边框 即只保留近似正方形的
x, y, w, h = r['rect']
if w / h > 1.2 or h / w > 1.2:
continue
candidates.add(r['rect'])
img_regions = []
for candidate in candidates:
x, y, w, h = candidate
img_regions.append(img[x:x + w][y:y + h])
return img_regions
def cropROI(filename):
global width, height
image = skimage.io.imread(filename)
width = len(image[0])
height = len(image)
# area = width * height
# if (area < 256*256 or area > 512*512) and ((height > 512 and width > 512) or (height < 256 and width <256)):
# height = int(512 * height/width)
# width = 512
num = 1
for sc in [350, 450, 500]:
for sig in [0.8]:
for mins in [30, 60, 120]:
img = skimage.io.imread(filename)[:, :, :3]
# if height == len(img) and width == len(img[0]):
# pass
# else:
# img = skimage.transform.resize(img, (height, width))
# perform selective search
img_lbl, regions = selectivesearch.selective_search(
img, scale=sc, sigma=sig, min_size=mins)
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 2000:
continue
# distorted rects
x, y, w, h = r['rect']
if w / h > 1.2 or h / w > 1.2:
continue
if w >= (img.shape[0] - 1) * (0.9) and h >= (img.shape[1] -
1) * (0.9):
continue
candidates.add(r['rect'])
print("Stage " + str(num) + " Done.")
num += 1
merge()
draw_superbox()
img = skimage.io.imread(filename)
# if height == len(img) and width == len(img[0]):
# pass
# else:
# img = skimage.transform.resize(img, (height, width))
i = 1
for x, y, w, h in refined:
if str(sys.version)[:3] == "3.4":
skimage.io.imsave(cmd_folder + "Temp/" + str(i) + ".jpg",
img[y:y + h, x:x + w]) # crop
else:
skimage.io.imsave(cmd_folder + "Temp/" + str(i) + ".jpg",
img[y:y + h, x:x + w]) # crop
i += 1
def image_proposal(img_path):
img = skimage.io.imread(img_path)
img_lbl, regions = selectivesearch.selective_search(img,
scale=500,
sigma=0.9,
min_size=10)
candidates = set()
images = []
vertices = []
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
if r['size'] < 220:
continue
# resize to 224 * 224 for input
proposal_img, proposal_vertice = prep.clip_pic(img, r['rect'])
# Delete Empty array
if len(proposal_img) == 0:
continue
# Ignore things contain 0 or not C contiguous array
x, y, w, h = r['rect']
if w == 0 or h == 0:
continue
# Check if any 0-dimension exist
[a, b, c] = np.shape(proposal_img)
if a == 0 or b == 0 or c == 0:
continue
im = Image.fromarray(proposal_img)
resized_proposal_img = resize_image(im, 224, 224)
candidates.add(r['rect'])
img_float = pil_to_nparray(resized_proposal_img)
images.append(img_float)
vertices.append(r['rect'])
return images, vertices
def my_selective_search(img):
img_array = np.array(img)
img_size = img_array.size / 3 #the number of img pixel
min_region_size = img_size / 100.0
min_component_size = int(math.sqrt(img_size) / 10)
#img_array = data.astronaut()
img_lbl, regions = select.selective_search(img_array,
scale=300,
sigma=0.9,
min_size=min_component_size)
#draw = ImageDraw.Draprint get_IOU([50, 50, 100 ,0], [0,0,100,100])w(img)
#print regions.__len__()
candidate = set()
n = 0
for i in regions:
if i['rect'] in candidate:
continue
if i['size'] < min_region_size:
continue
x, y, w, h = i['rect']
if w / h > 2 or h / w > 2:
continue
rect = (x, y, x + w, y + h)
candidate.add(rect)
return candidate
def getSelectiveSearchRois(img, ssScale, ssSigma, ssMinSize, maxDim):
# Selective Search
# Parameters
# ----------
# im_orig : ndarray
# Input image
# scale : int
# Free parameter. Higher means larger clusters in felzenszwalb segmentation.
# sigma : float
# Width of Gaussian kernel for felzenszwalb segmentation.
# min_size : int
# Minimum component size for felzenszwalb segmentation.
# Returns
# -------
# img : ndarray
# image with region label
# region label is stored in the 4th value of each pixel [r,g,b,(region)]
# regions : array of dict
# [
# {
# 'rect': (left, top, right, bottom),
# 'labels': [...]
# },
# ...
# ]
# inter_area seems to give much better results esp when upscaling image
img, scale = imresizeMaxDim(img, maxDim, boUpscale=True, interpolation = cv2.INTER_AREA)
_, ssRois = selectivesearch.selective_search(img, scale=ssScale, sigma=ssSigma, min_size=ssMinSize)
rects = []
for ssRoi in ssRois:
x, y, w, h = ssRoi['rect']
rects.append([x,y,x+w,y+h])
return rects, img, scale
def find_rois_selective_search(image_pixels, scale=200, sigma=0.9, min_size=10):
"""
Uses the selective search library to find rois
:param
image_path: path to an image
image_pixels: pixels from the image
"""
# Higher scale means higher preference for larger components (k / |C|, where |C| is the
# number of pixels in the component and k is the scale; for a large k, it would be difficult
# for a small component to be have a separation boundary with the neighboring components since
# the division is large). Smaller components are allowed when there is a sufficiently large
# difference between neighboring components (the higher k / |C|, the higher the difference
# between neighboring components has to be)
img_lbl, regions = \
selectivesearch.selective_search(image_pixels, scale=scale, sigma=sigma, min_size=min_size)
unique_rois = {}
# Deduplicating rois
for region in regions:
# rect format: [x, y, w, h]
rect = region["rect"]
key = str(rect[0]) + str(rect[1]) + str(rect[2]) + str(rect[3])
if key not in unique_rois:
# From [x, y, w, h] to {x, y, w, h}
unique_rois[key] = rect
return np.array(unique_rois.values())
def find_potential_regions(self, img, img_path):
# A crucial part of the program. It performs the selective search rather than a sliding window
# search. All areas that are found are stored in the regions structure.
img_lbl, regions = selectivesearch.selective_search(img,
scale=500,
sigma=0.5,
min_size=20)
# Used for filtering only.
candidates = []
# Keep any region of proper size
# Consider to remove detections for the full picture
coordinates = set()
for r in regions:
if (r['rect'] in coordinates):
continue
if (r['size'] < 2000):
continue
x, y, w, h = r['rect']
if (w / h > 1.2 or h / w > 1.2):
continue
print("Added somehting: ", x, " ", y, " ", x + w, " ", y + h)
coordinates.add(r['rect'])
candidates.append((img, np.array([[x, y, x + w, y + h]])))
# Last, evaluate if they are vehicles or somehting else
detected_vehicles = self.evaluate_detections(img, candidates)
# Then return the found vehicles
return detected_vehicles
def getSelectiveSearchRois(img, ssScale, ssSigma, ssMinSize, maxDim):
# Selective Search
# Parameters
# ----------
# im_orig : ndarray
# Input image
# scale : int
# Free parameter. Higher means larger clusters in felzenszwalb segmentation.
# sigma : float
# Width of Gaussian kernel for felzenszwalb segmentation.
# min_size : int
# Minimum component size for felzenszwalb segmentation.
# Returns
# -------
# img : ndarray
# image with region label
# region label is stored in the 4th value of each pixel [r,g,b,(region)]
# regions : array of dict
# [
# {
# 'rect': (left, top, right, bottom),
# 'labels': [...]
# },
# ...
# ]
# inter_area seems to give much better results esp when upscaling image
img, scale = imresizeMaxDim(img, maxDim, boUpscale=True, interpolation = cv2.INTER_AREA)
_, ssRois = selectivesearch.selective_search(img, scale=ssScale, sigma=ssSigma, min_size=ssMinSize)
rects = []
for ssRoi in ssRois:
x, y, w, h = ssRoi['rect']
rects.append([x,y,x+w,y+h])
return rects, img, scale
def image_proposal(img_path): img = skimage.io.imread(img_path) img_lbl, regions = selectivesearch.selective_search(img, scale=15, sigma=0.5, min_size=500) candidates = set() images = [] vertices = [] for r in regions: if r['rect'] in candidates: continue proposal_img, proposal_vertice = prep.clip_pic(img, r['rect']) if len(proposal_img) == 0: continue x, y, w, h = r['rect'] if w == 0 or h == 0: continue [a, b, c] = np.shape(proposal_img) if a == 0 or b == 0 or c == 0: continue im = Image.fromarray(proposal_img) array_img = Image.fromarray(proposal_img) resize_img = resize_image(array_img, 227, 227) img_float = pil_to_nparray(resize_img) images.append(img_float) vertices.append(r['rect']) return images, vertices
def main():
# loading lena image
# img = skimage.data.chelsea() # => numpy.ndarray
# print img[0][0]
# 書きの方法はダメ！ネガポジ画像になってしまう。
# img = Image.open("images/fruit.jpg")
# img.load()
# img = np.asarray( img, dtype="int32" )
# predictでは、data/bbox_image/n07747607_631_box2.JPEGは正解できた。
# ということは、n07747607_631.jpg でselective search すると正解できるかも？
tgt_img_path = "data/image/n07747607_631.jpg"
img = io.imread(tgt_img_path)
# perform selective search
img_lbl, regions = selectivesearch.selective_search(img, scale=500, sigma=0.9, min_size=10)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 2000:
continue
# distorted rects
x, y, w, h = r['rect']
if w / h > 1.2 or h / w > 1.2:
continue
candidates.add(r['rect'])
img = Image.open(tgt_img_path)
# imgs = [resize(img.crop(rect)) for rect in candidates]
res = []
for rect in candidates:
print rect
cropped_img = resize(img.crop(rect))
(answer, prob) = predict.predict_by_data(cropped_img)
if prob > 0.9:
res.append((rect,answer))
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(img)
for v in res:
x, y, w, h = v[0] # rect object
print v[1] # answer
print x, y, w, h
rect = mpatches.Rectangle((x, y), w, h, fill=False, edgecolor='red', linewidth=1)
ax.add_patch(rect)
plt.show()
def main():
# loading lena image
#img = skimage.data.lena()
imagefilename = 'TextImage.jpg'
img = np.float32(PIL.Image.open(imagefilename))/256.0
#showarray(img)
# perform selective search
img_lbl, regions = selectivesearch.selective_search(
img, scale=500, sigma=0.1, min_size=2)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 2000:
continue
# distorted rects
x, y, w, h = r['rect']
if w / h > 1.2 or h / w > 1.2:
continue
candidates.add(r['rect'])
# draw rectangles on the original image
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(img)
for x, y, w, h in candidates:
print x, y, w, h
rect = mpatches.Rectangle(
(x, y), w, h, fill=False, edgecolor='red', linewidth=1)
ax.add_patch(rect)
# create separate cropped images
i = 0
for x, y, w, h in candidates:
with Image(filename=imagefilename) as img:
print x, y, w, h
i = i + 1
oldimg = img
print img.size
img.crop(x,y,width=w,height=h)
img.format = 'jpeg'
img.save(filename='frame_'+str(i)+'.jpg')
img = oldimg
plt.show()
def generate_labels(self):
with codecs.open(self.all_list, 'r', 'utf-8') as f:
lines = f.readlines()
for num, image_idx in enumerate(lines):
ground_truth_dic = self.load_annotation(image_idx)
image_path = os.path.join(self.images_path, image_idx.strip() + '.jpg')
img = cv2.imread(image_path)
img_lbl, regions = selectivesearch.selective_search(img, scale=1000, sigma=0.9, min_size=1000)
labels = []
for r in regions:
x, y, w, h = r['rect']
proposal_vertice = [x + 1, y, x + w, y + h, w, h]
proposal_bbox = [x, y, (x + w - 1), (y + h - 1)]
label = np.zeros(self.class_num * 5 - 4, dtype=np.float32) # 假设包括背景有5类，0：5是判断类别，5：5+4*4=21 是位置框信息
iou_val=0
for ground_truth, class_idx in ground_truth_dic.items():
#ground_truth = list(ground_truth)
xmin=(2*ground_truth[0]-ground_truth[2])/2.0
ymin=(2*ground_truth[1]-ground_truth[3])/2.0
ground_truth=[xmin,ymin,ground_truth[2],ground_truth[3]]
iou_val = IOU(ground_truth, proposal_bbox)
px = float(proposal_vertice[0]) + float(proposal_vertice[4] / 2.0) # 中心点X
py = float(proposal_vertice[1]) + float(proposal_vertice[5] / 2.0) # 中心点Y
pw = float(proposal_vertice[4]) # w
ph = float(proposal_vertice[5]) # h
gx = float(ground_truth[0]) # 中心点X
gy = float(ground_truth[1]) # 中心点Y
gw = float(ground_truth[2]) # W
gh = float(ground_truth[3]) # H
if iou_val < self.roi_threshold :
label[0] = 1
elif iou_val > self.roi_threshold:
label[0] = 0
label[class_idx] = 1
label[self.class_num + (class_idx-1)*4 : self.class_num + (class_idx-1)*4 + 4] = \
[((gx - px) / pw), ((gy - py) / ph), (np.log(gw / pw)), (np.log(gh / ph))]
break
for i in range(len(proposal_bbox)):
proposal_bbox[i] = (proposal_bbox[i] / 16.0)
proposal_bbox.insert(0, 0)
proposal_bbox.insert(0, iou_val)
proposal_bbox.extend(label)
labels.append(proposal_bbox)
view_bar("Process image of %s" % image_path, num + 1, len(lines))
if self.is_save:
if not os.path.exists(self.processed_path): os.makedirs(self.processed_path)
np.save((os.path.join(self.processed_path, image_idx.split('.')[0].strip())
+ '_data.npy'), labels)
def main():
csv = open("bboxes.csv", "w")
csv.write("filename,ymin,xmin,ymax,xmax\n")
images = glob.iglob('test/*.png')
for i in images:
# loading lena image
img = skimage.io.imread(i)
# img = skimage.data.astronaut()
# perform selective search
img_lbl, regions = selectivesearch.selective_search(
img, scale=500, sigma=0.9, min_size=10)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 20 pixels
if r['size'] < 20:
continue
# distorted rects
x, y, w, h = r['rect']
if x is 0 or y is 0 or w is 0 or h is 0:
continue
if w / h > 1.5 or h / w > 1.5:
continue
candidates.add(r['rect'])
# draw rectangles on the original image
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(img)
for x, y, w, h in candidates:
#print x, y, w, h
csv.write('"nums/' + str(i) + '.png",' + str(y) + "," + str(x) + \
"," + str(y + h) + "," + str(x + w) + "\n")
rect = mpatches.Rectangle(
(x, y), w, h, fill=False, edgecolor='red', linewidth=1)
ax.add_patch(rect)
plt.show()
#plt.savefig("nums/" + str(i) + "boxes.png")
plt.close()
def selective_window(img,scale=300,sigma=0.8,mins=100):
#this is a time bottleneck
im_lables,regions = selective_search(img,scale=scale,sigma=sigma,min_size=mins);
true_regions = set([])
print '# of regions: {}'.format(len(regions))
for reg in regions:
if reg['rect'] in true_regions:
continue
if reg['size'] < 100:
continue
(x,y,w,h) = reg['rect']
if not (w*h > 0 and w/h < 4 and h/w < 4):
continue
true_regions.add(reg['rect'])
return true_regions
def selective(image_path):
# loading lena image
image_array = pre_process(image_path)
# perform selective search
# img = selectivesearch._generate_segments(
# image_array,
# scale=SELECTIVESEARCH_SCALE,
# sigma=SELECTIVESEARCH_SIGMA,
# min_size=SELECTIVESEARCH_MIN_SIZE
# )
# print('seg')
# print(img)
img_lbl, regions = selectivesearch.selective_search(
image_array,
scale=SELECTIVESEARCH_SCALE,
sigma=SELECTIVESEARCH_SIGMA,
min_size=SELECTIVESEARCH_MIN_SIZE
)
candidates = set()
for r in regions:
print('size')
x, y, w, h = r['rect']
print(r['size'])
print(w*h)
print(r['size'] == w*h)
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
# print('size')
# print(r['size'])
if r['size'] < 15*15:
continue
# distorted rects
x, y, w, h = r['rect']
if w / h > 3 or h / w > 3:
continue
candidates.add(r['rect'])
return post_process(candidates)
def main():
# loading lena image
img = skimage.data.chelsea() # => numpy.ndarray
print img[0][0]
# 書きの方法はダメ！ネガポジ画像になってしまう。
# img = Image.open("images/fruit.jpg")
# img.load()
# img = np.asarray( img, dtype="int32" )
# 参考：http://www.scipy-lectures.org/packages/scikit-image/
img = io.imread('images/fruit.jpg')
print img[0][0]
# perform selective search
img_lbl, regions = selectivesearch.selective_search(
img, scale=500, sigma=0.9, min_size=10)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 2000:
continue
# distorted rects
x, y, w, h = r['rect']
if w / h > 1.2 or h / w > 1.2:
continue
candidates.add(r['rect'])
# draw rectangles on the original image
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(img)
for x, y, w, h in candidates:
print x, y, w, h
rect = mpatches.Rectangle(
(x, y), w, h, fill=False, edgecolor='red', linewidth=1)
ax.add_patch(rect)
plt.show()
def process(net, image_folder, classes):
for image_path in sorted(glob.glob(image_folder+"/*.jpg")):
print 'Processing frame {}'.format(image_path)
im = cv2.imread(image_path)
# Detect all object classes and regress object bounds
timer = Timer()
timer.tic()
img_lbl, regions = selectivesearch.selective_search(im, scale=500, sigma=0.8, min_size=200)
regions_array = np.zeros((len(regions),4))
for idx in xrange(0,len(regions)):
rect = regions[idx]['rect']
regions_array[idx,:] = [rect[0],rect[1], rect[0]+rect[2], rect[1] + rect[3]]
obj_proposals = regions_array
timer.toc()
print ('Proposals took {:.3f}s for '
'{:d} object proposals').format(timer.total_time, len(regions))
timer = Timer()
timer.tic()
scores, boxes = im_detect(net, im, obj_proposals)
timer.toc()
print ('Detection took {:.3f}s for '
'{:d} object proposals').format(timer.total_time, boxes.shape[0])
# Visualize detections for each class
CONF_THRESH = 0.6
NMS_THRESH = 0.3
for cls in classes:
cls_ind = CLASSES.index(cls)
cls_boxes = boxes[:, 4*cls_ind:4*(cls_ind + 1)]
cls_scores = scores[:, cls_ind]
keep = np.where(cls_scores >= CONF_THRESH)[0]
cls_boxes = cls_boxes[keep, :]
cls_scores = cls_scores[keep]
dets = np.hstack((cls_boxes,
cls_scores[:, np.newaxis])).astype(np.float32)
keep = nms(dets, NMS_THRESH)
dets = dets[keep, :]
print 'All {} detections with p({} | box) >= {:.1f}'.format(cls, cls,
CONF_THRESH)
vis_detections(im, cls, dets, image_path, thresh=CONF_THRESH)
def main():
# loading lena image
#img = skimage.data.lena()
#print "oh hi"
img = np.float32(PIL.Image.open('sky1024px.jpg'))/256.0
#print "sup brah?"
# perform selective search
img_lbl, regions = selectivesearch.selective_search(
img, scale=500, sigma=0.95, min_size=50)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 2000:
continue
# distorted rects
x, y, w, h = r['rect']
if w / h > 1.2 or h / w > 1.2:
continue
candidates.add(r['rect'])
# draw rectangles on the original image
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(img)
for x, y, w, h in candidates:
print x, y, w, h
rect = mpatches.Rectangle(
(x, y), w, h, fill=False, edgecolor='red', linewidth=1)
ax.add_patch(rect)
plt.show()
def get_object_proposals(img, scale=500, sigma=0.9, min_size=10):
# Selective search
img_lbl, regions = selectivesearch.selective_search(
img, scale=scale, sigma=sigma, min_size=min_size)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 500 pixels
x, y, w, h = r['rect']
if r['size'] < 2000 or w > 0.95 * img.shape[1] or h > 0.95 * img.shape[0]:
continue
# excluding the zero-width or zero-height box
if r['rect'][2] == 0 or r['rect'][3] == 0:
continue
# distorted rects
if w / h > 5 or h / w > 5:
continue
candidates.add(r['rect'])
return candidates
def main():
# loading lena image
img = skimage.data.lena()
# perform selective search
img_lbl, regions = selectivesearch.selective_search(
img, scale=500, sigma=0.9, min_size=10)
print dir(regions)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# excluding regions smaller than 2000 pixels
if r['size'] < 2000:
continue
# distorted rects
x, y, w, h = r['rect']
if w / h > 1.2 or h / w > 1.2:
continue
candidates.add(r['rect'])
# save all candidates
i = 0
for c in candidates:
print i
x, y, w, h = c # x, y, width, height
print x, y, w, h
cropped = img[x:x+w, y:y+h]
fname = 'regions/region_{}.jpg'.format(i)
i += 1
skimage.io.imsave(fname, cropped)
if argc != 2:
print_usage(argvs[0])
quit()
# コマンドライン引数でファイルパスを指定
filepath = argvs[1]
# 画像読込
img = skimage.io.imread(filepath)
img = skimage.transform.resize(img, (64,64), mode='edge')
print('aaaaaaaaaa')
# selective search実行
label, regions = selectivesearch.selective_search(img, scale=64, sigma=0.8, min_size=2)
print('bbbbbbbbbb')
# 元画像にselective searchで取得した短形を追加して表示
candidates = set()
for reg in regions:
if reg['rect'] in candidates:
continue
if reg['size'] < 100:
continue
x, y, w, h = reg['rect']
#if float(w)*h > float(256)*256*0.95:
def load_2flowers(self):
with codecs.open(self.fineturn_list, 'r', 'utf-8') as f:
lines = f.readlines()
for num, line in enumerate(lines):
labels = []
labels_bbox = []
images = []
context = line.strip().split(' ')
image_path = context[0]
ref_rect = context[2].split(',')
ground_truth = [int(i) for i in ref_rect]
img = cv2.imread(image_path)
img_lbl, regions = selectivesearch.selective_search(img, scale=500, sigma=0.9, min_size=10)
candidate = set()
for r in regions:
if r['rect'] in candidate:
continue
if r['size'] < 200 :
continue
if (r['rect'][2] * r['rect'][3]) <500:
continue
proposal_img, proposal_vertice = clip_pic(img, r['rect'])
if len(proposal_img) == 0:
continue
x, y, w, h = r['rect']
if w==0 or h==0 :
continue
[a, b, c] =np.shape(proposal_img)
if a==0 or b==0 or c==0 :
continue
resized_proposal_img = resize_image(proposal_img, self.image_size,self.image_size)
candidate.add(r['rect'])
img_float = np.asarray(resized_proposal_img, dtype="float32")
if self.is_svm:
feature = self.solver.predict([img_float])
images.append(feature[0])
else :
images.append(img_float)
iou_val = IOU(ground_truth, proposal_vertice)
px = float(proposal_vertice[0]) + float(proposal_vertice[4] / 2.0)
py = float(proposal_vertice[1]) + float(proposal_vertice[5] / 2.0)
ph = float(proposal_vertice[5])
pw = float(proposal_vertice[4])
gx = float(ref_rect[0])
gy = float(ref_rect[1])
gw = float(ref_rect[2])
gh = float(ref_rect[3])
index = int(context[1])
if self.is_svm:
if iou_val < self.svm_threshold:
labels.append(0)
else:
labels.append(index)
label = np.zeros(5)
label[1:5] = [(gx - px) / pw, (gy - py) / ph, np.log(gw / pw), np.log(gh / ph)]
if iou_val < self.reg_threshold:
label[0] = 0
else:
label[0] = 1
labels_bbox.append(label)
else:
label = np.zeros(self.F_class_num )
if iou_val < self.fineturn_threshold :
label[0] = 1
else:
label[index] = 1
labels.append(label)
view_bar("Process SVM_and_Reg_image of %s" % image_path, num + 1, len(lines))
if self.is_save:
if self.is_svm:
if not os.path.exists(os.path.join(self.SVM_and_Reg_save_path, str(context[1]))):
os.makedirs(os.path.join(self.SVM_and_Reg_save_path, str(context[1])))
np.save((os.path.join(self.SVM_and_Reg_save_path, str(context[1]), context[0].split('/')[-1].split('.')[0].strip())
+ '_data.npy'),[images, labels, labels_bbox])
else:
np.save((os.path.join(self.fineturn_save_path, context[0].split('/')[-1].split('.')[0].strip()) +
'_data.npy'),[images, labels])
def get_ss_crops(img):
orig_img_height, orig_img_width, channels = img.shape
#print 'orig: ', str(orig_img_width), str(orig_img_height)
img = misc.imresize(img, (IMG_SIZE, IMG_SIZE))
img_height, img_width, channels = img.shape
#print float(img.size)
#print 'max size', float(img.size)/3*0.8
#print 'min size', float(img.size)/3*0.005
img_exp = img
#img_exp = exposure.equalize_adapthist(img, clip_limit=0.03)
# perform selective search
img_lbl, regions = selectivesearch.selective_search(
img_exp, scale=SCALE, sigma=SIGMA, min_size=MIN_SIZE)
candidates = set()
for r in regions:
# excluding same rectangle (with different segments)
if r['rect'] in candidates:
continue
# distorted rects
x, y, w, h = r['rect']
if h == 0 or w == 0:
continue
if w / h > 2 or h / w > 2:
continue
if w > img_width*0.9 or h > img_height*0.9:
continue
rect_size = w*h
#print rect_size
#print x, y, w, h
if rect_size > float(img.size)/3*0.8:
#print 'too big'
continue
if rect_size < float(img.size)/3*0.005:
#print 'too small'
continue
center_x = x + w/2
center_y = y + h/2
if center_x < img_width*0.15 or center_x > img_width*0.85 or center_y < img_height*0.15 or center_y > img_height*0.85:
#print 'off edges'
continue
# non-max suppression
candidates = [s for s in candidates if not (s[0] >= x or s[1] >= y or s[2] <= w or s[3] <= h)]
too_small = False
for larger in candidates:
if larger[0] <= x and larger[1] <= y and larger[2] >= w and larger[3] >= h:
#print x, y, w, h, " too small"
too_small = True
if too_small:
continue
new_width = int(w*WINDOW_EXPAND)
new_height = int(h*WINDOW_EXPAND)
x = max(0, center_x-new_width/2)
y = max(0, center_y-new_height/2)
w = min(img_width-x, new_width)
h = min(img_height-y, new_height)
#print x, y, w, h
x_scale_factor = float(orig_img_width)/IMG_SIZE
y_scale_factor = float(orig_img_height)/IMG_SIZE
#print x_scale_factor, y_scale_factor
x_scaled = x * x_scale_factor
y_scaled = y * y_scale_factor
w_scaled = w * x_scale_factor
h_scaled = h * y_scale_factor
candidates.append((x_scaled, y_scaled, w_scaled, h_scaled))
return candidates
def get_search(file_name, main_dir, sub_dir):
dir_path = main_dir + sub_dir
file_path = dir_path + '/' + file_name
open_img = Image.open('%s' % file_path)
img = np.array(open_img)
cropped_dir = crop_dir + sub_dir + '_crp'
# if file has already been processed, skip to next
if os.path.isfile(cropped_dir + '/' + file_name.strip('.jpg') + '.png'):
print "%s already processed, moving to next one..." % (sub_dir + '/' + file_name)
return
print "Searching for a whale in %s" % file_path
# perform selective search
img_lbl, regions = selectivesearch.selective_search(
img, scale=500, sigma=0.9, min_size=10)
print "Regions found: ", len(regions)
print "Selective search done. Now finding largest area of interest."
candidates = set()
rect_size = 0
# finds largest rectangle of interest - presumed to be whale
if len(regions) > 1:
for r in regions:
if r['rect'] in candidates:
continue
# excluding regions smaller than 40k px, or larger than 400k px
if r['size'] < rect_size or r['size'] < 40000 or r['size'] > 400000:
continue
# distorted rects
x, y, w, h = r['rect']
if w / h > 1.4 or h / w > 1.4:
continue
rect_size = r['size']
img_rect = r['rect']
if rect_size > 0:
print "Largest Size: ", rect_size
print "Rectangle = ", img_rect
make_cropped_img(open_img, img_rect, file_name, cropped_dir)
else:
# try again under less stringent conditions
for r in regions:
if r['rect'] in candidates:
continue
# excluding regions smaller than 20000 pixels
if r['size'] < rect_size or r['size'] < 20000 or r['size'] > 600000:
continue
# distorted rects
x, y, w, h = r['rect']
if w / h > 2.2 or h / w > 2.2:
continue
rect_size = r['size']
img_rect = r['rect']
if rect_size > 0:
print "Largest Size: ", rect_size
print "Rectangle = ", img_rect
print "Second pass successful..."
make_cropped_img(open_img, img_rect, file_name, cropped_dir)
else:
print "No potential whales found..."
text=set() text_cut = set() no_text = set() stage = 1 text_cut_final = set() for sc in [350,450,500]: for sig in [0.8]: for mins in [30,60,120]: # important img = skimage.io.imread(fname)[:,:,:3] if height == len(img) and width == len(img[0]): pass else: img = skimage.transform.resize(img, (height, width)) img_lbl, regions = selectivesearch.selective_search( img, scale=sc, sigma= sig,min_size = mins) for r in regions: # excluding same rectangle (with different segments) if r['rect'] in candidates: continue # excluding regions smaller than 2000 pixels if r['size'] < 2000: continue # distorted rects x, y, w, h = r['rect'] if w / h > 1.2 or h / w > 1.2: continue if w >= (img.shape[0]-1)*(0.7) and h >= (img.shape[1]-1)*(0.7): continue candidates.add(r['rect'])
from IPython.display import clear_output, Image, display
from google.protobuf import text_format
import caffe
import skimage.data
import selectivesearch
def showarray(a, fmt='jpeg'):
a = np.uint8(np.clip(a, 0, 255))
f = StringIO()
PIL.Image.fromarray(a).show() #.save(f, fmt)
img = skimage.data.lena()
img_lbl, regions = selectivesearch.selective_search(img, scale=500, sigma=0.9, min_size=10)
showarray(regions)
'''
regions
=> [{'labels': [0.0], 'rect': (0, 0, 59, 511), 'size': 15633},
{'labels': [1.0], 'rect': (58, 0, 1, 132), 'size': 173},
{'labels': [2.0], 'rect': (60, 18, 1, 5), 'size': 8},
{'labels': [3.0], 'rect': (5, 57, 25, 335), 'size': 507},
{'labels': [4.0], 'rect': (27, 166, 0, 9), 'size': 10},
{'labels': [5.0], 'rect': (59, 498, 0, 0), 'size': 1},
{'labels': [6.0], 'rect': (35, 484, 0, 27), 'size': 28},
{'labels': [7.0], 'rect': (58, 122, 4, 374), 'size': 581},
{'labels': [8.0], 'rect': (59, 40, 0, 81), 'size': 82},