超越百度的口罩检测算法
思百达物联网科技（北京）有限公司算法团队开源了一款轻量级cpu版实时口罩检测项目。
https://blog.csdn.net/jacke121/article/details/79846442
https://blog.csdn.net/weixin_37818440/article/details/104519213
https://99baiduyun.com/file/1nsQf_Py5YyKm87-8HiyJeQ.html
开源地址：
https://github.com/sbdcv/sbd_mask
人脸检测网络是centerface，是一款基于anchor free的高效的轻量级人脸检测网络。
centerface widerface测试集的精度：
| easy set | medium set | hard set |
|---|---|---|
| 93.2% | 92.1% | 87.3% |
口罩分类网络训练时使用自有数据集，MobileNetv2算法，官方声称准确率:
| 训练集 | 测试集 |
|---|---|
| 99% | 98.5% |
这是一款“速度和精度远超百度开源的轻量级口罩检测算法“
项目中给出了百度paddlehub的口罩检测算法预测代码，百度paddlehub口罩检测算法是有facebox开发的。
也给出了与百度paddlehub口罩检测算法的对比效果
测试速度在百度的两倍以上，精度也超过paddlehub口罩检测算法
下面是demo代码：
import cv2
import time
import numpy as np
from deploy import classify
from lib.CenterFace.centerface import CenterFace
def draw(bboxs, img=None, thresh=0.5, max_size=0):
img_cp = img.copy()
len_line = int(img_cp.shape[1] / 5)
pad_percent = int(img_cp.shape[1] / 2)
x = int(img_cp.shape[1] / 25)
y = int(img_cp.shape[0] / 25)
pad_x = int(img_cp.shape[1] / 50)
pad_y = int(img_cp.shape[0] / 25)
pad_text = 5
font_scale = (img_cp.shape[0] * img_cp.shape[1]) / (750 * 750)
font_scale = max(font_scale, 0.25)
font_scale = min(font_scale, 0.75)
font_thickness = 1
if max(img_cp.shape[0], img_cp.shape[1]) > 750: font_thickness = 2
if bboxs.shape[0] == 0: return img
bboxs = bboxs[np.where(bboxs[:, -1] > thresh)[0]]
bboxs = bboxs.astype(int)
cnt_mask = 0
cnt_nomask = 0
for bbox in bboxs:
img_bbox = img[bbox[1]:bbox[3], bbox[0]:bbox[2]]
if img_bbox.shape[0] * img_bbox.shape[1] < max_size:
continue
(ftype, prob) = classify(img_arr=img_bbox)
prob_font_scale = (img_bbox.shape[0] * img_bbox.shape[1]) / (100 * 100)
prob_font_scale = max(prob_font_scale, 0.25)
prob_font_scale = min(prob_font_scale, 0.75)
cv2.putText(img_cp, '{0:.2f}'.format(prob), (bbox[0] + 7, bbox[1] - 3),
cv2.FONT_HERSHEY_SIMPLEX, prob_font_scale, (0, 0, 255), 1, lineType=cv2.LINE_AA)
if ftype == 0: cnt_mask += 1
else: cnt_nomask += 1
color = (0, 0, 255) if ftype else (0, 255, 0)
cv2.rectangle(img_cp, (bbox[0], bbox[1]), (bbox[2], bbox[3]), color, 2)
cv2.line(img_cp, (x, y), (x + len_line, y), (0, 255, 0), 2)
cv2.putText(img_cp, 'Mask', (x + len_line + pad_x, y + pad_text),
cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 255, 0), font_thickness, lineType=cv2.LINE_AA)
cv2.line(img_cp, (x, y + pad_y), (x + len_line, y + pad_y), (0, 0, 255), 2)
cv2.putText(img_cp, 'face', (x + len_line + pad_x, y + pad_y + pad_text),
cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 255), font_thickness, lineType=cv2.LINE_AA)
mask_percent = (0 if cnt_mask == 0 else (cnt_mask / (cnt_mask + cnt_nomask))) * 100
cv2.putText(img_cp, 'Mask percent: {:.0f}%'.format(mask_percent), (x + pad_percent, y + pad_text),
cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 0), font_thickness, lineType=cv2.LINE_AA)
return img_cp
def detect(centerface, height=360, width=640, video_path='data/0.mp4', visualize=False):
cap = cv2.VideoCapture(video_path)
while (cap.isOpened()):
ret, frame = cap.read()
if ret == False: break
if frame is not None and frame.shape[0] > 0 and frame.shape[1] > 0:
frame = cv2.resize(frame, (width, height))
start_time = time.time()
dets, lms = centerface(frame, threshold=0.5)
frame = draw(dets, img=frame)
print("FPS: ", 1.0 / (time.time() - start_time))
if visualize:
max_size = 1024
if max(frame.shape[0], frame.shape[1]) > max_size:
scale = max_size / max(frame.shape[0], frame.shape[1])
frame = cv2.resize(frame, None, fx=scale, fy=scale)
cv2.imshow('sbd', frame)
# cv2.waitKeyEx()
if cv2.waitKey() & 0xFF == ord('q'):
break
cap.release()
cv2.destroyAllWindows()
if __name__ == "__main__":
im_width = 640
im_height = 360
net = CenterFace(im_height, im_width)
print('Finished loading model!')
detect(net, im_height, im_width, visualize=True)算法就是我们研发的，欢迎有需要合作的与我联系。
补充：最新又出来一个开源人脸口罩检测：
https://github.com/AIZOOTech/FaceMaskDetection
测试效果没有我们做的好：下图是随便举了两个图片例子，漏检和误检比较明显。
