def text_detect(img, prob=0.05):
im = Image.fromarray(img)
scale = IMGSIZE[0]
w, h = im.size
w_, h_ = resize_im(w, h, scale=scale,
max_scale=2048) ##短边固定为608,长边max_scale<4000
#boxed_image,f = letterbox_image(im, (w_,h_))
boxed_image = im.resize((w_, h_), Image.BICUBIC)
image_data = np.array(boxed_image, dtype='float32')
image_data /= 255.
image_data = np.expand_dims(image_data, 0) # Add batch dimension.
imgShape = np.array([[h, w]])
inputShape = np.array([[h_, w_]])
global graph
with graph.as_default():
##定义 graph变量 解决web.py 相关报错问题
"""
pred = textModel.predict_on_batch([image_data,imgShape,inputShape])
box,scores = pred[:,:4],pred[:,-1]
"""
box, scores = sess.run(
[box_score],
feed_dict={
textModel.input: image_data,
input_shape: [h_, w_],
image_shape: [h, w],
K.learning_phase(): 0
})[0]
keep = np.where(scores > prob)
box[:, 0:4][box[:, 0:4] < 0] = 0
box[:, 0][box[:, 0] >= w] = w - 1
box[:, 1][box[:, 1] >= h] = h - 1
box[:, 2][box[:, 2] >= w] = w - 1
box[:, 3][box[:, 3] >= h] = h - 1
box = box[keep[0]]
scores = scores[keep[0]]
return box, scores
def show(p, scale=608):
im = Image.open(p)
xmlP = p.replace('.jpg', '.xml').replace('.png', '.xml')
boxes = read_voc_xml(xmlP)
im = Image.open(p)
w, h = resize_im(im.size[0], im.size[1], scale=scale, max_scale=4096)
input_shape = (h, w)
isRoate = True
rorateDegree = np.random.uniform(-5, 5)
rorateDegree = 0
newBoxes, newIm = get_box_spilt(boxes,
im,
sizeW=w,
SizeH=h,
splitW=splitW,
isRoate=isRoate,
rorateDegree=rorateDegree)
return plot_boxes(im, 0, boxes,
color=(0, 0, 0)), plot_box(newIm, newBoxes), newBoxes
def get_xml_box_wh(self, filename):
xmlP = filename.replace('.jpg', '.xml').replace('.png', '.xml')
boxes = read_voc_xml(xmlP)
im = Image.open(filename)
scale = np.random.choice(self.scales, 1)[0]
w, h = resize_im(im.size[0], im.size[1], scale=scale, max_scale=2048)
input_shape = (h, w)
isRoate = False
rorateDegree = 0
newBoxes, newIm = get_box_spilt(boxes,
im,
sizeW=w,
SizeH=h,
splitW=self.splitW,
isRoate=isRoate,
rorateDegree=rorateDegree)
box = []
for bx in newBoxes:
w = int(bx[2] - bx[0])
h = int(bx[3] - bx[1])
box.append([w, h])
return box
def data_generator(roots, anchors, num_classes, splitW):
'''
data generator for fit_generator
@@roots:jpg/png
'''
n = len(roots)
np.random.shuffle(roots)
scales = [416, 608, 608, 608] ##多尺度训练
i = 0
j = 0
m = len(scales)
while True:
root = roots[i]
i += 1
if i >= n:
i = 0
scale = scales[j]
j += 1
if j >= m:
j = 0
xmlP = os.path.splitext(root)[0] + '.xml'
boxes = read_voc_xml(xmlP)
im = Image.open(root)
w, h = resize_im(im.size[0], im.size[1], scale=scale, max_scale=None)
if max(w, h) > 2048:
w, h = resize_im(im.size[0],
im.size[1],
scale=scale,
max_scale=2048)
input_shape = (h, w)
isRoate = True
rorateDegree = np.random.uniform(-5, 5)
rorateDegree = 0
newBoxes, newIm = get_box_spilt(boxes,
im,
w,
h,
splitW=splitW,
isRoate=isRoate,
rorateDegree=rorateDegree)
newBoxes = np.array(newBoxes)
if len(newBoxes) == 0:
continue
if np.random.randint(0, 100) > 70:
if np.random.randint(0, 100) > 50:
##图像水平翻转
newBoxes[:, [0, 2]] = w - newBoxes[:, [2, 0]]
im = Image.fromarray(cv2.flip(np.array(im), 1))
else:
##垂直翻转
newBoxes[:, [1, 3]] = h - newBoxes[:, [3, 1]]
im = Image.fromarray(cv2.flip(np.array(im), 0))
maxN = 128 ##随机选取128个box用于训练
image_data = []
box_data = []
image, box = get_random_data(newIm,
newBoxes,
input_shape,
max_boxes=maxN)
image_data = np.array([image])
box_data = np.array([box])
y_true = preprocess_true_boxes(box_data, input_shape, anchors,
num_classes)
yield [image_data, *y_true], [np.zeros(1)] * 4
