Research And Implementation Of Object Detection In Underwater Images Based On Faster R-CNN

The recent years have witnessed a sharp population growth worldwide.Increasingly close attention,therefore,has been paid to the exploitation and utilization of ocean resources.Currently,Underwater operations are mainly carried out by manual labor,which has the disadvantages of high risk,low efficiency and high cost.Hence,it is of great significance to develop an underwater fishing robot that can automatically detect,locate and fish underwater objects.Object detection play an important role in the work of underwater fishing robots,and its results will directly affect the follow-up planning and control of machinery.The existing object detection algorithms mainly focus on the object detection tasks of land images.However,due to the problems of image degradation,object clustering and occlusion,and difficulty in large-scale accurate labeling,underwater images cannot obtain higher detection accuracy only by the existing methods.Consequently,this paper focuses on the object detection in underwater images and uses the Faster R-CNN object detection framework as the benchmark framework.Specifically,it includes the following contents:Firstly,the main causes of underwater image degradation are analyzed,and the underwater style migration data set is constructed by combining two traditional image enhancement methods.On this basis,an underwater image enhancement method based on Generative Adversarial Networks is proposed.This method achieves the purpose of image enhancement by employing the idea of style migration,and realizes color correction and detail restoration of underwater images.Secondly,a weakly supervised object detection method based on underwater foreground segmentation is proposed to solve the problem that data sets are difficult to label accurately due to the accumulation of objects.The method consists of an underwater foreground segmentation stage based on U-Net and a object detection stage based on underwater foreground segmentation.The first stage is used to complete the segmentation of underwater foreground and background.The second stage completes the object detection task based on the first stage.In this paper,two strategies are proposed for the second stage,namely,the mean filling strategy and the proposal-refined strategy.The proposed method eliminates the false negative problem caused by incomplete annotation,and it can obtain higher detection accuracy with incomplete annotated datasetThirdly,the problem of few-shot caused by difficult acquisition of underwater images is studied.Through the idea of transfer learning,the knowledge learned from the marine classification network is transferred to the object detection network.Meantime,the spatial transformation network is used to augment the data of underwater images,which improves the robustness of the model to spatial transformation while expanding the data.Based on the above research contents,this paper designs and implements underwater object detection system.A large number of qualitative and quantitative experiments have been carried out to verify the effectiveness of the proposed method.The experimental results show that the method proposed in this paper can still obtain high detection accuracy even if the image degradation is serious and the number of datasets is not enough and annotation is not accurate enough.The proposed method improved the performance of object detection in underwater scene.It is of great significance to promote the intelligence of marine capture machines.