Research On Feature Enhancement Methods Of Detection Network Based On Deep Learning

The deep learning based detection networks have been widely used in various computer vision tasks.Enhancing the feature of detection network can further improve its performance and application performance.In recent years,the feature enhancement for detection networks has become the current research hot-spot,and a large number of methods have emerged.However,as the variability of detection network architecture and the diversity of application fields,the current methods usually focus on enhancing the feature of a particular detection network architecture,while the general and easy-to-use feature enhancement methods for detection networks still need to be further explored and studied.To address the contradiction between spatial feature and semantic feature that are both common in detection networks,this thesis focus on the enhancement methods based on feature aggregation,so as to achieve the transmission of cross-level semantic features and spatial features in detection networks,further strengthen expression ability of feature in the detection network,to enhance the overall detection performance of network.The main works and contributions of this thesis are summarized as follows:(1)A nonlinear aggregation based feature enhancement method for detection networks is proposed.For the detection networks,the higher-level feature layers which have strong semantic features while weak spatial features,that hinder the accurate prediction of networks.In order to balance the semantic and spatial features in detection networks,this thesis proposed a nonlinear aggregation based feature enhancement method,which fuses feature layers in adjacent stages by iterative ways,and enhances the expression ability of the feature layers in the network.The effectiveness of this method has been verified on the tasks of both object detection and semantic segmentation.(2)A multistage stacking feature enhancement method is proposed that is general for detection networks.For a classic encoder-decoder architecture,the decoder is susceptible to the interference of complex background due to the local receptive fields of the convolution kernels during the decoding process.This problem further results more false positive predictions.To salve this problem,a multistage feature stacking method isproposed.This method first stores global feature through feature aggregation,and then redistributes that in the decoding process,which greatly avoids the interference of complex background feature during the decoding process.The effectiveness of this method has been verified on the tasks of object detection and semantic segmentation.(3)A semantic segmentation architecture with network-level feature enhancement is proposed.The introduction of feature enhancement methods in the detection networks will bring additional parameters and computational effort.Therefore,this thesis focus on a network-level feature enhancement method,which is embodied in a semantic segmentation architecture.The architecture completes one-step decoding and up-sampling by aggregating the feature layers of the decoder at all stages,which greatly reduces the network parameters and improves the train-ability of the network.When the number of training samples is reduced to 1/4 of the original training samples,the segmentation accuracy is still 1.4% higher than that of the trained benchmark network on the complete training samples.