Efficient And Lightweight Feature Pyramid Network For Object Detection

Object detection is one of the most challenging problems in computer vision.Its purpose is to detect the position of specific objects in images.It has been widely used in face recognition,automatic driving,pedestrian detection,video surveillance and other fields.Deep learning based methods has become the mainstream method with the capability of superior feature representation,automatic feature extraction and high-precision detection.In this paper,we propose the multi-feature concatenation network(MFCN)and the lightweight feature pyramid network(LFPN).The main work is as follows:Compared with the image pyramid based method,the feature pyramid-based method is faster and able to make full advantages of the feature information from different convolution layers.The existing feature pyramid-based method fuses feature maps of different scales by adding corresponding elements,result in losing some low-level detail feature information frequently.To tackle the problem mentioned above,this paper proposes a multi-feature concatenation network based on the feature pyramid network(FPN).A concatenating structure of multi-layer feature maps is designed to achieve feature fusion between different semantic layers,which can effectively surmount the loss of low-level detailed information.At the same time,combined with Res Net101,multi-scale object detection is realized.Experimental results on benchmark datasets show that the proposed algorithm is superior to FPN.Compared with the one-stage approach,the two-stage approach can achieve higher detection accuracy.In the two-stage detection framework,the classification regression subnetwork adopts two sets of fully connected layer designs,resulting in slow detection speed.To solve above problem,this paper proposes a lightweight feature pyramid network.By replacing the traditional fully connected layer with deep separable convolution,and the lightweight classified regression subnetwork is designed.At the same time,the feature map compression module is designed to reduce the number of feature maps to the region proposal network.Experimental results on benchmark datasets show that the proposed algorithm far exceeds FPN with less accuracy loss.