Algorithm Research And Implementation For Lightweight Network Target Detection In Specific Application Scenarios

Synthetic aperture radar(SAR)is a microwave sensor that uses electromagnetic wave scattering characteristics for imaging.It has certain capabilities of penetrating clouds and ground,and can achieve all-weather observation.It has unique advantages in ocean monitoring,surveying and mapping,military and other fields.With the continuous development of marine resources,people have begun to pay attention to the use of SAR for detecting marine ship targets.Due to the special imaging mechanism and wide monitoring range of SAR,there are problems such as unclear imaging ship contours,large scale differences,and dense arrangement of small ships.At the same time,in order to adapt to the application scenarios of SAR ship target detection algorithms on portable mobile terminals,it is also necessary to solve the problem of algorithm lightweight.These problems have become one of the current research hotspots in this field.This thesis focuses on the above problems and completes the following research around object detection algorithms in deep learning:(1)An Multi-scale SAR image detection algorithm for ships based on improved YOLOv5 is proposed to address the large pixel scale difference of ship targets in complex scenes and missed detection caused by dense array of ships.For the neck network of YOLOv5,a bi-directional feature pyramid network(BiFPN)is adopted to enhance the multi-scale feature fusion ability of the network,and an EC-MLP(Enhanced channel-MLP)module is constructed based on depthwise separable convolution(DSC)and channel MLP in its bottom-up feature fusion branch to enrich semantic information and provide more sufficient ship target context features.The global attention mechanism(GAM)is introduced to enable the network to extract input features selectively and reduce information reduction.In addition,the SIoU loss function is used to further improve the training convergence speed and detection accuracy of the network.Comparative experiments with eight other methods(FasterR-CNN,Libra R-CNN,FCOS,YOLOv5s,PP-YOLOv2,YOLOX-s,PP-YOLOE-s and YOLOv7-tiny)are conducted on SSDD and HRSID datasets.The experimental results show that the AP50 of the improved algorithm reaches 96.7%on SSDD and 95.6%on HRSID,which is superior to the comparison methods.(2)A lightweight SAR ship detection algorithm based on improved YOLOX is proposed to solve the problems of limited hardware computing resources in micro synthetic aperture radar(SAR)platforms and complex backgrounds,large differences in ship scales,and dense distribution of small targets in SAR images.Firstly,based on depthwise separable convolution(DSC)and Ghost module,the backbone network in YOLOX is designed for lightweight processing.SepViT(Separable Vision Transformer)module is used to solve the feature channel information separation caused by DSC and enhance global information interaction of the network.Secondly,for feature fusion network,a duplicate bilateral feature fusion network is designed based on PANet(Path Aggregation Network),which obtains richer semantic information by repeatedly extracting and fusing features.Finally,adaptively spatial feature fusion(ASFF)is added to the prediction network to adaptively learn the weights of different scale features,and fully merge the semantic information of high-level features and the fine-grained information of low-level features.(3)In response to the YOLOv5 algorithm,channel dimension model pruning is performed,and the Gamma parameter of the batch normalization layer(BN)is introduced to perform sparse training with L1 norm.The channels with activation levels less than the threshold are pruned to remove redundant connections and parameters in the network model to reduce the number of network parameters,thereby reducing computational complexity and memory consumption.