Research And Implementation Of Marine Organism Object Detection System Based On Deep Learning

Marine organism observation is an important topic in the field of marine object detection,providing an effective way to assess the abundance of marine organisms and changes in the marine ecological environment.Monitoring holothurians,scallops,and other seafood in marine ranches on a large scale and continuously for a long period of time by underwater robotic platforms can not only help farmers to control the growth status and quantity changes of marine organisms in real-time and avoid working in a high-risk environment but also help to analyze the changes of marine ecological environment.However,underwater object detection faces many challenges,such as underwater image blurring,object scale variation,and background complexity,which are not conducive to model learning and reduce the performance of underwater object detection.Therefore,in this thesis,we research and implement a marine organism object detection system by improving the YOLOv5 algorithm.The details of the research are as follows:(1)In the dataset,the underwater images show severe green color bias and low contrast due to light absorption and scattering.In order not to affect the model’s full learning of object features,this thesis uses the RGHS image enhancement algorithm to correct the contrast and color to improve the quality of underwater images.For the enhanced images,an improved Mosaic-9 data enhancement is used to enrich the small object and improve the generalization performance of the model.(2)Due to the small scale of the image object,the features are easily lost with the deepening of the convolution layer.Therefore,the process of 4-fold downsampling is added to the original YOLOv5 structure,and a prediction head for detecting small objects is added based on three YOLO Head branches,while shallow features are fused with high-resolution semantic information,which is conducive to better learning of features by the model,enhancing the sensitivity of the model to small target detection and improving the degree of model miss detection.(3)Because of the complex marine environment,false detection occurs when foreground objects are intertwined with the background.Therefore,a triplet attention mechanism is introduced in the YOLOv5 network structure to enhance the attention weights of the focus region,improve the model’s ability to extract local features with discriminative power,suppress the background useless information,and improve the degree of model misdetection.(4)The research and implementation for the marine organism object detection system based on deep learning with the improved YOLOv5 algorithm model according to the practical needs of this thesis.Develop a visual interface to facilitate user learning and use.Experiments and tests show that the system has the usability and stability to meet practical application requirements.