基于改进的Cascade R-CNN算法的低光照人脸检测方案

Low-Light Face Detection Scheme Based on Improved Cascade R-CNN Algorithm

对低光照条件下图像中的人脸进行检测是当前人脸识别及其衍生技术中十分重要且具有挑战性的任务。该任务的难度在于低光照条件下摄像头能见度低、信号细节损失严重、目标检测难。针对上述问题,本文提出一种改进的Cascade R-CNN算法,以提升低光照条件下人脸检测精度。首先,使用K-Means算法对数据中的真实标注框进行聚类,从而对Cascade R-CNN中的锚框尺寸进行修正,其次,在Cascade R-CNN网络中添加增加反卷积层提升小目标识别精度;最后使用多种训练技巧进行组合,提升识别精度。实验结果表明,经过改进的Cascade R-CNN模型,与使用原始的Cascade R-CNN算法进行检测相比mAP提升了36.8%。

Face detection in images under low illumination is a very important and challenging task in face recognition and its derivatives.The difficulty of this task lies in the low visibility of the camera,the serious loss of signal details and the difficulty of target detection under low light conditions.To solve the above problems,this paper proposes an improved cascade r-cnn algorithm to improve the accuracy of face detection under low light conditions.Firstly,the k-means algorithm is used to cluster the real annotation boxes in the data,so as to correct the anchor box size in cascade r-cnn.Secondly,an anti convolution layer is added to cascade r-cnn network to improve the accuracy of small target recognition;Finally,a variety of training techniques are combined to improve the recognition accuracy.The experimental results show that the improved cascade r-cnn model improves the map by 36.8%compared with the original cascade r-cnn algorithm.