Multi-scale Target Detection Based On Feature Fusion

Due to the influence of the shooting distance,the size of the target within the same shooting range changes greatly,resulting in missed detection of the detector,which is an important challenge for the deep learning-based target detection method.Therefore,improving the precision of multi-scale target detection is a key issue in the current target detection application research.The birth of FPN network greatly improves the performance of existing deep learning target detection algorithm for multi-scale target detection,however,the single feature fusion path of FPN network cannot guarantee the full fusion of features at different layers of convolutional neural network,and will lead to the dilution of deep features in the fusion process,resulting in the limited association between deep features and shallow features established through feature fusion.To this,on the basis of FPN network,this thesis proposes a feature from deep to shallow overlay fusion method,the method based on convolution between neural network and the FPN network,let the characteristics of the deep integration of convolution neural network,the more effective in relieving the characteristics of deep dilution problem in the process of integration,enhanced the correlation characteristic of different layers.In order to ensure sufficient feature fusion,this thesis added the method after FPN network,and carried out experiments based on two-stage target detection model.In the target detection experiment of Pascal VOC dataset,the m AP of the proposed model is 80.6%,which is 2.4% improved compared with the target detection model of FPN network.In order to verify the performance of the proposed model for multiscale target detection,a separate test was conducted on pedestrian categories in Pascal VOC datasets,and the AP reached 81.3%,2.8% higher than the target detection model of FPN network.Image features extracted by convolutional neural network are saved and transmitted by channels.Deep feature maps usually have more channels and express more specific information,which are generally local features of a certain region.These local features may be the features of the region where the target is located or the features of irrelevant target detection tasks.Paying more attention to the targetrelated channel information in the feature fusion process can optimize the feature fusion process and improve the performance of multi-scale target detection.Therefore,based on FPN network and feature superposition fusion method from deep to shallow,this thesis proposes a method to guide the deep feature and shallow feature fusion of convolutional neural network through channel attention network,so that the network pays more attention to the channel containing target information in the process of feature fusion.However,FPN network will adjust the number of feature channels of different layers to 256,which will cause information loss for the deepest features,affect the performance of channel attention network,and is not conducive to target detection.To solve this problem,this thesis adds a module to alleviate the information loss of the deepest features caused by the adjustment of the number of channels.The experimental results on Pascal VOC dataset show that the performance of the module alleviating information loss for target detection is improved.The performance of the information loss suppression module for object detection is improved.On this basis,the model m AP after adding channel attention network is improved compared with that before.It is proved that the method of superposition fusion of features guided by channel attention network from deep to shallow can effectively improve the performance of multi-scale target detection.