Research And Implementation Of Underwater Occluded Object Detection Method Based On GAN

Due to the complexity of the underwater environment,object detection in underwater scenes is more challenging than land and air target detection.Underwater creatures' occlusion and overlap between aquatic creatures due to their living habits and external sediment factors,as well as the low contrast,blurred details,and color deviation of underwater images,are the current difficulties in underwater object detection.Aiming at the problems of low detection accuracy and missed detection of existing object detection methods in underwater occluded objects,This paper proposes a fusion of GAN and Faster R-CNN underwater occlusion object detection model to improve the detection accuracy in occlusion and overlapping scenes,and reduce the number of missed objects.The main research work of this thesis is as follows:1.To deal with the problem of blurred details and color degradation of underwater images in the data set,image preprocessing is performed to obtain images with clear details and color correction.In addition,in view of the insufficient amount of occluded samples in the data set,the data amount of occluded samples is expanded by random occlusion to provide sufficient data for model training.2.Propose an optimized generative confrontation network,which increases the quality of generated images,alleviates the phenomenon of gradient disappearance in the network transmission process,introduces residual blocks in the generative confrontation network,and performs feature fusion through jump connections to improve the overall network The stability and keep the image information transmitted in the network.3.Improve Faster R-CNN,use ROI-Align to map the feature map in the candidate area to a fixed-size feature map,further improve the detection accuracy of the network;not only that,this article will also use the original Faster R-CNN The maximum suppression algorithm(NMS)is replaced with a soft non-maximum suppression algorithm(Soft-NMS),which reduces the number of missed targets in occlusion and overlapping situations.4.Combining the improved Faster R-CNN with the optimized generative confrontation network,an underwater occlusion target model based on the generative confrontation network is proposed.The optimized generative confrontation network is used to generate the features of the occlusion image,and the generated features are used as Improve the input of Faster R-CNN to further improve the model's low detection accuracy and missed detection under occlusion and overlap.