Marine Ranching Biological Object Detection Research Based On Deep Learning

Marine ranching biological object detection is a key technology for developing and utilizing marine resources.Rapid and accurate detection of marine ranching organisms is of great importance for the sustainable development and protection of marine resources.In the traditional process of marine ranching biological detection,features often need to be manually designed and completed by machine learning.However,the underwater environment is complex,and it is difficult for researchers to design features based on their past experiences.Their robustness and portability are also poor,making it difficult to meet the actual needs of engineering.The research and development of computer vision in recent years has been rapid,and deep learning has been widely used in trajectory tracking,object tracking,and object detection.This has greatly improved the accuracy of detection in these areas.This thesis aims to overcome the limitations of traditional marine ranching biological detection by designing a deep learning-based marine ranching object detection system,which was applied to the underwater biological detection in the Luhaifeng Marine Ranch.The main contents are as follows:1.This thesis analyzes the problem of insufficient clarity in underwater images to be tested,mainly including underwater image haze,blur,color contrast,and blurred object contours and texture features.Therefore,appropriate image enhancement algorithms are needed to preprocess the images for testing.The effects of four image enhancement algorithms,including white balance,histogram equalization,adaptive histogram equalization with limited contrast,and dark channel prior,are compared.The algorithm’s performance is evaluated using the SSIM(Structural Similarity)and PSNR(Peak Signal to Noise Ratio)parameters,and the adaptive histogram equalization with limited contrast is selected as the image enhancement algorithm after comprehensive evaluation.Image data augmentation is used to expand the training dataset in order to imporve the performance and ability of the model to generalize.2.The basic framework of the object detection model in this thesis uses the YOLO(You Only Look Once)V3 algorithm.The principle and limitations of YOLO V3 are analyzed.To further improve the accuracy of the detection algorithm,the Kmeans++ clustering algorithm is used to obtain anchors that are more adaptable to the model.The effects of IoU(Intersection over Union),GIoU(Generalized Intersection over Union),and DIoU(Distance Intersection over Union)as regression loss functions are compared,and DIoU is selected to replace the original regression loss function in YOLO V3.To address the problem of the large size of the DarkNet53 backbone network in YOLO V3,this thesis uses EfficientNet-B2 to replace DarkNet53 as the backbone network for the object detection algorithm.This reduces the size of the model,enabling it to work normally in environments with limited computing resources.3.In order to meet the practical needs of marine ranching object detection,this thesis built a marine ranching object detection system based on Py Qt in the Pycharm development environment,which includes a data management module,an image preprocessing interface,and a object detection interface.The system is easy to operate and effectively improves the efficiency of marine ranching object detection.Finally,different sets of seven marine organism images were used to test the system’s detection performance,and the results showed that the detection performance was good and could meet practical needs.