def face_detection_image(net, image_name):
imgs = cv2.imread(image_name)
rows, cols, ch = imgs.shape
scales = caculate_scales(imgs)
total_boxes = []
for scale in scales:
w, h = int(rows * scale), int(cols * scale)
scale_img = tf.resize(imgs, (w, h))
# scale_img = cv2.resize(imgs,(w,h))/255.0
net.blobs["data"].reshape(1, channel, w, h)
transformer = caffe.io.Transformer({"data": net.blobs["data"].data.shape})
transformer.set_transpose("data", (2, 0, 1))
transformer.set_channel_swap("data", (2, 1, 0))
transformer.set_raw_scale("data", raw_scale)
out = net.forward_all(data=np.asarray([transformer.preprocess("data", scale_img)]))
boxes = generateBoundingBox(out["prob"][0, map_idx], scale)
if boxes:
total_boxes.extend(boxes)
boxes_nms = np.array(total_boxes)
true_boxes = nms_max(boxes_nms, overlapThresh=0.3)
true_boxes = nms_average(np.array(true_boxes), overlapThresh=0.07)
img_cv = render_result.read_image(image_name)
result = img_cv
for box in true_boxes:
result = render_result.draw_rectangle(result, (int(box[0]), int(box[1]), int(box[2]), int(box[3])), (0, 0, 255))
render_result.save_2_file(result, "result/" + image_name.split("/")[-1])
return true_boxes
def face_detection_image(net,image_name):
imgs = cv2.imread(image_name)
rows,cols,ch = imgs.shape
scales = caculate_scales(imgs)
total_boxes = []
for scale in scales:
w,h = int(rows* scale),int(cols* scale)
scale_img = tf.resize(imgs,(w,h))
#scale_img = cv2.resize(imgs,(w,h))/255.0
net.blobs['data'].reshape(1,channel,w,h)
transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
transformer.set_transpose('data', (2,0,1))
transformer.set_channel_swap('data', (2,1,0))
transformer.set_raw_scale('data', raw_scale)
out = net.forward_all(data=np.asarray([transformer.preprocess('data', scale_img)]))
boxes = generateBoundingBox(out['prob'][0,map_idx], scale)
if(boxes):
total_boxes.extend(boxes)
boxes_nms = np.array(total_boxes)
true_boxes = nms_max(boxes_nms, overlapThresh=0.3)
true_boxes = nms_average(np.array(true_boxes), overlapThresh=0.07)
img_cv = render_result.read_image(image_name)
result = img_cv
for box in true_boxes:
result = render_result.draw_rectangle(result,(int(box[0]),int(box[1]),int(box[2]),int(box[3])),(0,0,255))
render_result.save_2_file(result,'result/'+image_name.split('/')[-1])
return true_boxes
def face_detection_image(net, net_vf, image_name):
'''
@检测单张人脸图像
'''
scales = []
imgs = skimage.io.imread(image_name)
if imgs.ndim == 3:
rows, cols, ch = imgs.shape
else:
rows, cols = imgs.shape
#计算需要的检测的尺度因子
min = rows if rows <= cols else cols
max = rows if rows >= cols else cols
# 放大的尺度
delim = 2500 / max
while (delim >= 1):
scales.append(delim)
delim = delim - 0.5
#缩小的尺度
min = min * factor
factor_count = 1
while (min >= face_w):
scale = pow(factor, factor_count)
scales.append(scale)
min = min * factor
factor_count += 1
#=========================
#scales.append(1)
total_boxes = []
###显示热图用
num_scale = len(scales)
s1 = int(np.sqrt(num_scale)) + 1
tt = 1
plt.subplot(s1, s1 + 1, tt)
plt.axis('off')
plt.title("Input Image")
im = caffe.io.load_image(image_name)
plt.imshow(im)
#============
for scale in scales:
w, h = int(rows * scale), int(cols * scale)
scale_img = tf.resize(imgs, (w, h))
#更改网络输入data图像的大小
net.blobs['data'].reshape(1, channel, w, h)
#转换结构
transformer = caffe.io.Transformer(
{'data': net.blobs['data'].data.shape})
#transformer.set_mean('data', np.load(caffe_root + 'python/caffe/imagenet/ilsvrc_2012_mean.npy').mean(1).mean(1))
transformer.set_transpose('data', (2, 0, 1))
transformer.set_channel_swap('data', (2, 1, 0))
transformer.set_raw_scale('data', raw_scale)
#前馈一次
out = net.forward_all(
data=np.asarray([transformer.preprocess('data', scale_img)]))
###显示热图用
tt = tt + 1
plt.subplot(s1, s1 + 1, tt)
plt.axis('off')
plt.title("sacle: " + "%.2f" % scale)
plt.imshow(out['prob'][0, map_idx])
#===========
boxes = generateBoundingBox(out['prob'][0, map_idx], scale)
if (boxes):
total_boxes.extend(boxes)
#非极大值抑制
boxes_nms = np.array(total_boxes)
true_boxes1 = nms_max(boxes_nms, overlapThresh=0.3)
true_boxes = nms_average(np.array(true_boxes1), overlapThresh=0.07)
#===================
plt.savefig('heatmap/' + image_name.split('/')[-1])
#在图像中画出检测到的人脸框
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(imgs)
for box in true_boxes:
im_crop = im[box[0]:box[2], box[1]:box[3], :]
if im_crop.shape[0] == 0 or im_crop.shape[1] == 0:
continue
if re_verify(net_vf, im_crop) == True:
rect = mpatches.Rectangle((box[0], box[1]),
box[2] - box[0],
box[3] - box[1],
fill=False,
edgecolor='red',
linewidth=1)
ax.text(box[0],
box[1] + 20,
"{0:.3f}".format(box[4]),
color='white',
fontsize=6)
ax.add_patch(rect)
plt.savefig('result/' + image_name.split('/')[-1])
plt.close()
return out['prob'][0, map_idx]
def face_detection_image(net, net_vf, info):
'''
@检测单张人脸图像
'''
image_name = '/home/mjd/FaceDetection_CNN/aflw/data/' + info[1]
scales = []
imgs = skimage.io.imread(image_name)
if imgs.ndim == 3:
rows, cols, ch = imgs.shape
elif imgs.ndim == 0: #09437 channel is 0
return 0
else:
rows, cols = imgs.shape #grey image
imgs = skimage.color.gray2rgb(imgs)
#计算需要的检测的尺度因子
min = rows if rows <= cols else cols
max = rows if rows >= cols else cols
# 放大的尺度
delim = 2500 / max
while (delim >= 1):
scales.append(delim)
delim = delim - 0.5
#缩小的尺度
min = min * factor
factor_count = 1
while (min >= face_w):
scale = pow(factor, factor_count)
scales.append(scale)
min = min * factor
factor_count += 1
#=========================
#scales.append(1)
total_boxes = []
###显示热图用
num_scale = len(scales)
s1 = int(np.sqrt(num_scale)) + 1
tt = 1
plt.subplot(s1, s1 + 1, tt)
plt.axis('off')
plt.title("Input Image")
im = caffe.io.load_image(image_name)
plt.imshow(im)
#============
for scale in scales:
w, h = int(rows * scale), int(cols * scale)
scale_img = tf.resize(imgs, (w, h))
#更改网络输入data图像的大小
net.blobs['data'].reshape(1, channel, w, h)
#转换结构
transformer = caffe.io.Transformer(
{'data': net.blobs['data'].data.shape})
#transformer.set_mean('data', np.load(caffe_root + 'python/caffe/imagenet/ilsvrc_2012_mean.npy').mean(1).mean(1))
transformer.set_transpose('data', (2, 0, 1))
transformer.set_channel_swap('data', (2, 1, 0))
transformer.set_raw_scale('data', raw_scale)
#前馈一次
out = net.forward_all(
data=np.asarray([transformer.preprocess('data', scale_img)]))
###显示热图用
tt = tt + 1
plt.subplot(s1, s1 + 1, tt)
plt.axis('off')
plt.title("sacle: " + "%.2f" % scale)
heatmap = plt.imshow(out['prob'][0, map_idx])
plt.colorbar(heatmap) #添加颜色指示条
#===========
boxes = generateBoundingBox(out['prob'][0, map_idx], scale)
if (boxes):
total_boxes.extend(boxes)
#非极大值抑制
boxes_nms = np.array(total_boxes)
true_boxes1 = nms_max(boxes_nms,
overlapThresh=0.2) #paras according to experiment
true_boxes = nms_average(np.array(true_boxes1), overlapThresh=0.03)
#===================
plt.savefig('heatmap/' + image_name.split('/')[-1])
#在图像中画出检测到的人脸框
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(imgs)
#draw groundtruth rect
rect = mpatches.Rectangle((float(info[2]), float(info[3])),
float(info[4]),
float(info[5]),
fill=False,
edgecolor='blue',
linewidth=1)
ax.add_patch(rect)
#count correct detection boxes
result = []
box_num = 0
correct_num = 0
for box in true_boxes:
#check if detected rect matches groundtruth
xx1 = np.maximum(float(info[2]), box[0])
yy1 = np.maximum(float(info[3]), box[1])
xx2 = np.minimum(float(info[2]) + float(info[4]), box[2])
yy2 = np.minimum(float(info[3]) + float(info[5]), box[3])
w = np.maximum(0, xx2 - xx1)
h = np.maximum(0, yy2 - yy1)
area_box = (box[2] - box[0]) * (box[3] - box[1])
area_gt = float(info[4]) * float(info[5])
overlap = (w * h) / (area_gt + area_box - w * h)
box_num = box_num + 1
if overlap > 0.3:
correct_num = correct_num + 1
im_crop = im[box[0]:box[2], box[1]:box[3], :]
if im_crop.shape[0] == 0 or im_crop.shape[1] == 0:
continue
if True: #re_verify(net_vf, im_crop) == True:
rect = mpatches.Rectangle((box[0], box[1]),
box[2] - box[0],
box[3] - box[1],
fill=False,
edgecolor='red',
linewidth=1)
ax.text(box[0],
box[1] + 20,
"{0:.3f}".format(box[4]),
color='white',
fontsize=6)
ax.add_patch(rect)
plt.savefig('result/' + image_name.split('/')[-1])
plt.close()
result.append(box_num)
result.append(correct_num)
result.append(out['prob'][0, map_idx])
return result
def face_detection_image(net,net_vf,image_name):
'''
@检测单张人脸图像
'''
scales = []
imgs = skimage.io.imread(image_name)
if imgs.ndim==3:
#print imgs.shape
rows,cols,ch = imgs.shape
else:
rows,cols = imgs.shape
#计算需要的检测的尺度因子
min_this = rows if rows<=cols else cols
max_this = rows if rows>=cols else cols
# 放大的尺度
delim = 2500/max_this
while (delim >= 1):
scales.append(delim)
delim=delim-0.5
#缩小的尺度
min_this = min_this * factor
factor_count = 1
while(min_this >= face_w):
scale = pow(factor, factor_count)
scales.append(scale)
min_this = min_this * factor
factor_count += 1
#=========================
#scales.append(1)
total_boxes = []
###显示热图用
num_scale = len(scales)
s1=int(np.sqrt(num_scale))+1
tt=1
plt.subplot(s1, s1+1, tt)
plt.axis('off')
plt.title("Input Image")
im=caffe.io.load_image(image_name)
plt.imshow(im)
#============
for scale in scales:
try:
w,h = int(rows* scale),int(cols* scale)
scale_img= tf.resize(imgs,(w,h))
#print scale_img.shape
#更改网络输入data图像的大小
net.blobs['data'].reshape(1,channel,w,h)
#转换结构
transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
#transformer.set_mean('data', np.load(caffe_root + 'python/caffe/imagenet/ilsvrc_2012_mean.npy').mean(1).mean(1))
transformer.set_transpose('data', (2,0,1))
transformer.set_channel_swap('data', (2,1,0))
transformer.set_raw_scale('data', raw_scale)
#前馈一次
out = net.forward_all(data=np.asarray([transformer.preprocess('data', scale_img)]))
#print out['prob'][0,0].shape
###显示热图用
tt=tt+1
plt.subplot(s1, s1+1, tt)
plt.axis('off')
plt.title("sacle: "+ "%.2f" %scale)
plt.imshow(out['prob'][0,map_idx])
#print out['prob'][0,0]
#===========
boxes = generateBoundingBox(out['prob'][0,map_idx], scale)
if(boxes):
total_boxes.extend(boxes)
except:
continue
#非极大值抑制
boxes_nms = np.array(total_boxes)
true_boxes1 = nms_max(boxes_nms, overlapThresh=0.3)
true_boxes = nms_average(np.array(true_boxes1), overlapThresh=0.07)
#===================
plt.savefig('/home/test/tangyudi/faceDetection_alfw/heatmap/'+image_name.split('/')[-1])
#在图像中画出检测到的人脸框
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(imgs)
for box in true_boxes:
im_crop = im[box[0]:box[2],box[1]:box[3],:]
if im_crop.shape[0] == 0 or im_crop.shape[1] == 0:
continue
#if re_verify(net_vf, im_crop) == True:
rect = mpatches.Rectangle((box[0], box[1]), box[2]-box[0], box[3]-box[1],
fill=False, edgecolor='red', linewidth=1)
ax.text(box[0], box[1]+20,"{0:.3f}".format(box[4]),color='white', fontsize=6)
ax.add_patch(rect)
plt.savefig('/home/test/tangyudi/faceDetection_alfw/result/'+image_name.split('/')[-1])
plt.close()
return out['prob'][0,map_idx]
fig = plt.figure()
im2 = plt.imshow(out['prob'][0,1], cmap = plt.get_cmap('jet'))
pos = fig.add_axes([0.93, 0.21, 0.02, 0.56])
fig.colorbar(im2, cax=pos)
plt.savefig("Results\\Heatmaps\\" + str(name))'''
#bounding boxes around the predicted faces locations
bboxes = makeBoundingBoxes(out['prob'][0, 1], sf)
if (bboxes):
total_bboxes.extend(bboxes)
#the following scheme is : NMS_AVERAGE scheme. Perform greater than NMS_MAX scheme.
all_bboxes = np.array(total_bboxes)
preserved_bboxes = nms_max(all_bboxes, OVERLAPPING_THRESH)
preserved_bboxes = nms_average(np.array(preserved_bboxes), CLUSTER_SCORE_THRESH)
'''#the following scheme is : NMS_MAX scheme.
all_bboxes = np.array(total_bboxes)
preserved_bboxes = nms_max(all_bboxes, OVERLAPPING_THRESH)'''
#write image name onto opened .txt files
resultData.write(thisImgPath + " \n")
#using Draw method in Object ImageDraw
draw = ImageDraw.Draw(img)
'''###
box_counter = 0
temp = np.zeros(5)
def face_detection_image(net, net_vf, image_name):
'''
@检测单张人脸图像
'''
if _DEBUG != True:
pdb.set_trace()
scales = []
imgs = skimage.io.imread(image_name)
if imgs.ndim == 3:
rows, cols, ch = imgs.shape
elif imgs.ndim == 0: #09437 channel is 0
return 0
else:
rows, cols = imgs.shape #grey image
imgs = skimage.color.gray2rgb(imgs)
#计算需要的检测的尺度因子
min = rows if rows <= cols else cols
max = rows if rows >= cols else cols
# 放大的尺度
delim = 2500 / max
while (delim >= 1):
scales.append(delim)
delim = delim - 0.5
#缩小的尺度
min = min * factor
factor_count = 1
while (min >= face_w):
scale = pow(factor, factor_count)
scales.append(scale)
min = min * factor
factor_count += 1
#=========================
#scales.append(1)
total_boxes = []
###显示热图用
num_scale = len(scales)
s1 = int(np.sqrt(num_scale)) + 1
tt = 1
plt.subplot(s1, s1 + 1, tt)
plt.axis('off')
plt.title("Input Image")
im = caffe.io.load_image(image_name)
plt.imshow(im)
#============
for scale in scales:
w, h = int(rows * scale), int(cols * scale)
scale_img = tf.resize(imgs, (w, h))
#更改网络输入data图像的大小
net.blobs['data'].reshape(1, channel, w, h)
#转换结构
transformer = caffe.io.Transformer(
{'data': net.blobs['data'].data.shape})
#transformer.set_mean('data', np.load(caffe_root + 'python/caffe/imagenet/ilsvrc_2012_mean.npy').mean(1).mean(1))
transformer.set_transpose('data', (2, 0, 1))
# if imgs.ndim==3:#some pictures are in black and white
transformer.set_channel_swap('data', (2, 1, 0))
transformer.set_raw_scale('data', raw_scale)
#前馈一次
out = net.forward_all(
data=np.asarray([transformer.preprocess('data', scale_img)]))
###显示热图用
tt = tt + 1
plt.subplot(s1, s1 + 1, tt)
plt.axis('off')
plt.title("sacle: " + "%.2f" % scale)
heatmap = plt.imshow(out['prob'][0, map_idx])
plt.colorbar(heatmap) #添加颜色指示条
#===========
boxes = generateBoundingBox(out['prob'][0, map_idx], scale)
if (boxes):
total_boxes.extend(boxes)
#非极大值抑制
info = image_name.split('/')
info1 = info[8].split('.')
boxes_total = np.array(total_boxes)
draw_boxes(boxes_total, imgs, im, info1[0] + 'total.jpg')
#找出每个热图中最大的若干概率值
# idxs = np.argsort(boxes_total[:,4])
# last = len(idxs) - 1
# i = idxs[last]
#写入文件
#fileWriter = open( './result/max_probs.txt', 'a+' )
#fileWriter.write( image_name )
#for i in idxs[ last-10 : last ]:
#fileWriter.write( ' ' )
# fileWriter.write( str(np.asscalar(boxes_total[i,4])) )
#fileWriter.write( '\n' )
#fileWriter.close()
true_boxes_max = nms_max(boxes_total, 0.2)
#draw_boxes( true_boxes_max, imgs, im, info1[0]+'true_boxes_max.jpg')
#true_boxes_avg = nms_average(boxes_total, 0.2)
#draw_boxes( true_boxes_avg, imgs, im, info1[0]+'true_boxes_avg,jpg')
true_boxes_max_avg = nms_average(np.array(true_boxes_max), 0.05)
draw_boxes(true_boxes_max_avg, imgs, im, info1[0] + '_1.jpg')
#true_boxes_avg_max = nms_max(np.array(true_boxes_avg), 0.07)
#draw_boxes( true_boxes_avg_max, imgs, im, info1[0]+'true_boxes_avg_max.jpg')
true_boxes_max_avg = nms_average(np.array(true_boxes_max), 0.07)
draw_boxes(true_boxes_max_avg, imgs, im, info1[0] + '_2.jpg')
true_boxes_max_avg = nms_average(np.array(true_boxes_max), 0.1)
draw_boxes(true_boxes_max_avg, imgs, im, info1[0] + '_3.jpg')
true_boxes_max = nms_max(boxes_total, 0.3)
true_boxes_max_avg = nms_average(np.array(true_boxes_max), 0.05)
draw_boxes(true_boxes_max_avg, imgs, im, info1[0] + '_4.jpg')
true_boxes_max_avg = nms_average(np.array(true_boxes_max), 0.07)
draw_boxes(true_boxes_max_avg, imgs, im, info1[0] + '_5.jpg')
true_boxes_max_avg = nms_average(np.array(true_boxes_max), 0.1)
draw_boxes(true_boxes_max_avg, imgs, im, info1[0] + '_6.jpg')
true_boxes_max = nms_max(boxes_total, 0.4)
true_boxes_max_avg = nms_average(np.array(true_boxes_max), 0.05)
draw_boxes(true_boxes_max_avg, imgs, im, info1[0] + '_7.jpg')
true_boxes_max_avg = nms_average(np.array(true_boxes_max), 0.07)
draw_boxes(true_boxes_max_avg, imgs, im, info1[0] + '_8.jpg')
true_boxes_max_avg = nms_average(np.array(true_boxes_max), 0.1)
draw_boxes(true_boxes_max_avg, imgs, im, info1[0] + '_9.jpg')
#===================
#plt.savefig('heatmap/'+image_name.split('/')[-1])
#在图像中画出检测到的人脸框
plt.close()
return out['prob'][0, map_idx]
