Feature Enhancement Based Single Shot Multibox Detector

Object Detection is one of the most critical basic techniques in computer vision systems.For many artificial intelligence systems in the real world,accurate target positioning is essential,such as target tracking,identification and alignment.Thanks to the powerful feature extraction capabilities of convolutional neural networks,object detection algorithm has made breakthrough.However,in some applications which is high demand on time and space complexity of algorithms,such as drone navigation?security field and automatic driving,the existing real-time target detection algorithms cannot achieve satisfactory results in terms of detection performance.Therefore,how to improve the accuracy of real-time target detection network has gained wide attention.However,the existing improvements to the real-time target detection algorithm are mostly at the expense of time and space complexity.Aimed at this shortage,this paper designs two kinds of feature enhancement methods based on the state-of-the-art real-time object detection network SSD,which enhances the discriminability of convolutional neural network features form the perspective of space and channel?These two methods effectively improve the detection accuracy of SSD network with marginal extra time cost,and reduce a certain space complexity.The main work of this paper includes the following two aspects:Firstly,based on the SSD network,a spatial feature enhancement network is proposed based on the problem that surrounding background or unrelated target interference disturb the target feature extraction process.This network draws on the human visual attention mechanism and proposes a lightweight spatial attention module,which aims to make the network focus on detecting key areas in the scene at a small cost.The module uses the multi-scale characteristics of convolutional neural network features to generate attention maps through the deep feature layer of the network to constrain the learning of adjacent shallow feature layers.In the experiment,compared with the SSD network,the detection performance of the spatial feature enhancement network has a great advantage,and the difference in model complexity is small.Secondly,aiming at the problem of feature redundancy,this paper proposes a multi-scale channel feature enhancement method.This method proposes a channel feature enhancement network.The network design a channel feature enhancement module,which can makethe network suppress the learning of redundant features and enhance the discriminability between features.The channel feature enhancement module quantizes the importance of the feature on a channel-by-channel basis to obtain a feature importance vector that is used to suppress redundant features.This feature enhancement method has enabled the SSD network to achieve significant improvements in the public dataset Pascal VOC.Finally,this paper combines the above two feature enhancement methods to improve the network feature extraction ability from the perspective of space and channel,and proposes a multi-scale feature enhancement network.The experimental results show that the multiscale feature enhancement network has achieved similar accuracy with some of the most advanced real-time target detection networks,and has higher detection efficiency.