复杂背景下小尺寸多角度人脸检测方法研究

Research on small-scale,multi-angle face detection methods in complex backgrounds

为了提升复杂背景下小尺寸人脸检测精度,提出了一种人脸检测方法GhostNet-MTCNN。在多任务级联卷积神经网络(MTCNN)主干网络上,将占用计算资源的普通卷积进行舍弃,利用GhostNet网络中计算量更低的Ghost bottleneck模组替代卷积的作用,重新构建网络特征提取功能,从而搭建一个新的模型。实验结果表明,该方法可以有效平衡参数量和精度。在Easy、Medium、Hard三种验证集上,与MTCNN相比在参数量仅增加0.62M的前提下精度分别提升了5.6%、6.6%、7.8%,与MobileNetV3-MTCNN相比在参数量减少1.27M的同时精度又分别提升了1.6%、0.8%、0.5%。该研究能够在复杂场景下提高模型对小尺寸、多角度人脸检测精度,同时也能够有效平衡参数量和检测精度使其成为在边缘设备部署中更优的选择。

A face detection approach which is named GhostNet-MTCNN was proposed to enhance the precision of small sized face detection in complex backgrounds.On the backbone of MTCNN,this approach uses the lower computational Ghost bottleneck module which is in the GhostNet to replace the convolutional function,and discards the common convolution which occupies computer resources to configure the network′s feature extraction function.Through the process,a new module will be set up.The experimental results showed that the approach can effectively balance parameter quantity and precision.Across three validation sets categorized as Easy,Medium and Hard,compared to the original MTCNN,the proposed GhostNet-MTCNN achieves notable improvements in accuracy respectively 5.6%,6.6%and 7.8%,while the parameter quantity only with a minimal increase of 0.62M.Furthermore,compared to MobileNetV3-MTCNN,GhostNet-MTCNN outperforms by enhancing accuracy by 1.6%,0.8%and 0.5%,meanwhile a reduction in parameter quantity by 1.27M.The study can not only enhance the precision of the module to detect the small-sized and multi-angle faces in complex backgrounds but also can effectively balance parameter quantity and detection precision,which will make it a superior choice for edge deployment devices.