Research On Underwater Target Detection Algorithm Based On Improved SSD

Convolutional neural networks have become the mainstream method for object detection,but these existing methods are not fully applicable to underwater object detection.There are mainly problems such as blurred underwater images,small and large aggregations of underwater organisms,and underwater devices that are not suitable for using large depth neural networks.In order to solve the above problems,this thesis investigates the above problems using existing underwater datasets,and mainly accomplishes the following work:(1)To address the problems of image distortion in underwater dataset and more underwater small targets affecting the accuracy of underwater target detection,this thesis firstly constructs an image enhancement module CCNet to enhance the underwater images,the idea of the enhancement module is to color convert the underwater parti-color images,the color converted images are more conducive to the subsequent feature extraction and improve the generalization ability of the detection model;secondly in the detection module uses ResNeXt network with stronger feature extraction capability as the backbone network;finally,in the neck network,Improved Feature Pyramid Network(IFPN)is used to make shallow features obtain more semantic information of higher-level features and alleviate the problem that feature information of small targets will gradually decrease during the sampling process.The image enhancement module is jointly trained with the object detection module to alleviate the problem of model information loss in both stages and further improve the model detection capability.The detection accuracy of the algorithm reaches 78.4% on the UPRC dataset.(2)To address the problem that underwater storage devices cannot deploy large detection networks,this thesis improves the detection network with lightweight,and designs a lightweight underwater object detection model SG-Det.Firstly,a new lightweight feature extraction network SGNet is proposed,which greatly reduces the number of model parameters;secondly,a cross-scale feature fusion module(AFF)is designed,which introduces an attention mechanism The input features are weighted in both global and local channels to highlight useful information,thus reducing the interference of irrelevant information such as background;finally,the fourth layer with higher nonlinearization is selected to enhance the semantic information of the first three layers respectively by practical AFF module,so that the first three layers can perform better in recognizing small objects at a smaller cost.The algorithm achieves 71.75%accuracy on the dataset UPRC with only 5.43 M number of parameters.Figure [31] Table [12] Reference [79]...