Compression On Deep Face Detection Models Under Resource-limited Environment

Face detection technology is important for face-related tasks such as face alignment and face recognition and is widely used in real life.Existing face detection models can be divided into DNN-based face detection models and traditional face detection models.DNN-based face detection models are far more accurate than the traditional ones.However,DNN-based face detection models are difficult to be used in real-world applications.For example,MTCNN and Face Box are two different DNN-based models that can achieve real-time detection on personal computers.However,resource-limited devices such as embedded devices have a small storage space that these two models are difficult to be deployed in those devices.Therefore,these two models need to be compressed to reduce the number of parameters and computation cost.However,existing methods only study how to prune classification models while MTCNN and Face Box are detection models.The network structure of these two models is different from the classification model.Therefore,using the existing methods to prune MTCNN and Face Box directly will lead to a serious loss of performance.To solve this problem,we study the compression of MTCNN and Face Box according to their network structure.The main research contents of this paper can be summarized as follows.MTCNN uses a cascade structure to detect faces.However,existing methods directly set the same pruning rate on the cascade structure will result in a large performance loss.To solve this problem,we first propose a pruning strategy for MTCNN to reduce the performance loss of pruned MTCNN.The strategy not only considers the correlation between each model in MTCNN but also considers the redundancy of each model.At the same time,to further reduce the performance loss of pruned MTCNN,we also propose a fine-tuning method based on knowledge distillation.This method uses the original model to guide the fine-tuning process of the pruning model.At last,we propose a compression framework for the DNN-based face detection model by combining channel pruning and quantization.This framework uses low-bit quantization to significantly reduce the model size of MTCNN after pruning.Face Box is a one-stage DNN-based face detection model that has a different network structure from classification models.However,use existing methods to prune Face Box directly will result in a large performance loss.To solve this problem,we first propose a channel pruning method for one-stage DNN-based face detection models to reduce the performance loss of pruned Face Box.This method uses Face Box's output branches to assist the selection of channels with strong discriminating ability in its skeleton network.After that,to further reduce the performance loss of pruned Face Box,we propose a new fine-tuning method based on knowledge distillation according to the network structure of one-stage DNN-based face detection models.This method uses the outputs features of hidden layers in the original model and its outputs together to guide the fine-tuning process of the pruning model.Besides,we also use low-bit quantization to significantly reduce the model size of pruned Face Box.At last,experiments show that our proposed compression method can effectively reduce the calculation cost and model size of the DNN-based face detection model without causing significant performance degradation.