Research On Light Weight Face Detection Algorithm Based On Multi-scale Anchor Densification Strategy

The performance of face detection algorithms improves continuously with the rapid development of the convolutional neural networks.However,the current methods still can be difficult to balance between detection speed and accuracy,and light weight face detection algorithms have low recall rate for small and dense face target.The paper proposes a light weight face detection algorithm based on multi-scale anchor densification strategy aims at the problems above.The main work are as follows:Firstly,based on the SSD face detection framework,we design a lightweight face detection algorithm,which is mainly composed of rapid convolutional layer,Inception module and multi-scale convolutional layer.Rapid convolutional layer quickly reduces the size of input feature map by designing big stride and big kernel size convolutional layers,which can alleviate computational cost of network;Inception module obtain higher feature learning efficiency by making use of convolutional kernel with different size to process feature information parallelly;Multi-scale convolutional layer rationally redistributes the range of multi-scale features and design the size of Anchors according to the receptive fields of different features,which make sure all of the Anchors cover all kinds of face targets.Then the algorithm is further optimized based on the network structure introduced above.It mainly includes the following three items:(1)The rapid convolutional layer doubles the number of channels without computational cost and keep the negative phase information of network by CReLU activation function.Therefore,CReLU greatly decreases the time complexity of convolutional layers which have big stride and big kernel size while keep the accuracy of network.In the meantime,rapid convolutional layer replaces standard convolutional layers with depthwise separable convolution to greatly reduce time consumption of network but not decrease much detection accuracy.(2)At first,Inception module bring in Dilated Inception to simulate the function of eccentricity in human visual cortex,which enhance the feature discriminability and robustness of Inception module.And then Inception module bring in Squeeze and Excitation module to realize the channel-wise attention mechanism.and makes sure model can pay more attention to the relatively important information.Finally,we construct ADI module by combining the two part above.(3)Multi-scale convolutional layer increases the face detection result by bringing in Anchor densification strategy which enlarge the intersection area between Anchors.In addition,Max-out background label is introduced to alleviate the imbalanced between positive and negative samples and reduce the false positive rate of detection result.Finally,experiment result demonstrate that the face detection algorithm proposed by the paper has fast face detection speed,small model complexity and robustness against small and dense face target which increase face recall rate of light weight algorithm effectively.The experiment result on FDDB database show that the algorithm proposed by this paper has 92.91%face recall rate when the number of false positive is 2000,and the average accuracy rates on AFW and PASCAL FACE database are 97.21%and 90.21%respectively.The algorithm proposed by this paper has 4FPS face detection speed on the E5-2643v4@3.40 CPU and 100FPS face detection speed on the single TiTan X GPU.In addition,the physical memory of the model is only 4.1M.