Research On Fish Object Detection In Underwater Complex Environment

China is a large country of marine fisheries,and with the development of fisheries-related industries,it has become more and more important to our national economic income,among which the detection and identification of underwater fish have important research value in aquaculture,water monitoring,fisheries resources research,etc.Traditional methods of fish detection are often done manually by detecting and identifying fish in videos or images taken underwater;or by removing fish from the water to a well-lit laboratory environment for extraction and identification.This approach not only requires a lot of human and material resources,but also requires the personnel responsible for detection to have a certain knowledge base as well as experience in fish research,which is often very inefficient and does not achieve the expected results.The method of stripping the fish from the water body,although it can be better for feature extraction,loses the actual effect of its activity in the water and is difficult to use in real production environments.Nowadays,the practical application of computer vision technology in underwater fish videos is developing rapidly,and more and more algorithms are appearing instead of manual methods for object detection of fish.This type of object detection algorithm can be used to detect and identify targets for fish in real production environments using actual video or images taken of underwater fish activity.Not only does it not require a lot of manual operations,but it can also achieve unmonitored real-time detection.However,due to the actual production environment,the quality of the underwater video or image taken,affected by a variety of conditions such as the water environment,light irradiation,and filming equipment,often have a dark scene,blurred target fish,and more impurities in the water environment,resulting in poor filming results and difficult object detection tasks.To address the above problems,this paper proposes an underwater fish object detection method based on GMG background removal and PLS classifier for object detection of fish in underwater videos.And for the part in which the fish object detection in the night video is poor,an underwater fish object detection method that first enhances the original image using the MSRCP algorithm and then improves the framework based on the Cascade R-CNN model is proposed.The main work and results of this paper are as follows.(1)Two types of datasets,video and image,were produced by shooting underwater video data.Using an underwater camera,video data of fish activities in a pond were captured and divided into two different scenes,day and night,according to the light intensity.The total video duration is 270 minutes,and the file size is 18.1G of which100 minutes are daytime video,100 minutes are nighttime video,and 70 minutes are a mixture of daytime and nighttime video.At the same time,the captured videos are intercepted and converted into images according to a fixed number of frames,totaling10,800 images with a resolution of 720*270.9,000 of them are processed as the training set and 1,800 as the test set,and the fish in the images are annotated with Lab Image and made into the XML format required for the experiment.(2)In this paper,GMG background removal algorithm and PLS classifier are used for object detection of fish in underwater videos.First,the collected video dataset is processed into frame format and its default RGB format is converted to HSI format.The luminance part of the frame is denoised using the median filter method,and then the foreground and background of the current frame are separated using the GMG background removal method.In addition,local binary pattern(LBP)texture features as well as grayscale coefficient features are extracted and input to a partial least squares(PLS)classifier to detect fish in two different scenes,day and night,respectively,in this paper.Experimental results show that the accuracy of object detection of fish using this method is 96.89%.The object detection of fish in video can be performed quickly without manual detection.It further provides a theoretical basis for the development and utilization of underwater resources.(3)For the experimental results of the image scene taken underwater at night is cloudy and blurred,making the effect of post-object detection poor.In this paper,MSRCP image enhancement and the improved Cascade R-CNN model are used for object detection of fish in underwater images.Firstly,for underwater images at night,the MSRCP algorithm is used for image enhancement to improve the color and sharpness of the images,and then input to the Det NASNet backbone network for network training and feature information extraction,and the extracted feature information is input to the Cascade R-CNN model,and the Soft-NMS candidate frame optimization algorithm is used for which the RPN network is The object detection of fish at night under shady conditions was achieved by optimization.The experimental results show that the detection accuracy of underwater fish image object detection at night can reach 95.81% using this method,which is 11.57% better than the original Cascade R-CNN method.The method solves the problems of image degradation and overlapping fish object detection in nighttime underwater environment,and achieves fast detection of underwater fish targets at night.