Research On Marine Biological Target Detection Based On Deep Learning

China is a large marine area country,covering an area of up to 3 million square meters,with its rich marine resources.With the continuous growth of China’s population,land resources are no longer enough to meet people’s daily production and living needs,so more and more people are turning their attention to marine resources.In recent years,marine target detection technology has been widely used in the fields of seafood fishing,marine ecosystem health assessment,and marine biodiversity detection.Compared with traditional image detection,marine target detection is relatively difficult because natural light is scattered by suspended particles in water and absorbed by water during its propagation underwater,resulting in low clarity of underwater images acquired by humans.In addition,the scattering of light underwater tends to blur the details of underwater images,while the absorption of light by water usually leads to distortion of underwater image colors,resulting in serious blue-green color bias and reduced contrast in the underwater images captured by the camera.The underwater images are the basis of marine biological target detection,and the existence of the above problems makes the marine biological target detection face a major challenge.In addition,the existing detection methods have good accuracy,but the network model size is large and the computation speed is slow,which cannot meet the demand for real-time detection of marine organisms.To address these problems,this paper uses underwater image processing algorithms as the preprocessing part of the marine life target detection network and optimizes the marine life target detection network to design a deep learning-based marine life target detection system,which is divided into a network preprocessing part and a marine life target detection part.In the network preprocessing part,a multi-scale Retinex-based underwater image enhancement algorithm is designed,which obtains the reflected V component by transforming the color space of the underwater image,then estimates the V component using bilateral filtering as the surround function of the multi-scale Retinex algorithm,and finally uses CLAHE to contrast the reflected V component,the initial H component and the initial S component respectively The final enhanced underwater image is obtained by merging and converting to RGB color space after enhancement.Five image quality evaluation indexes are selected to compare and analyze the enhancement effect from both subjective and objective aspects,to verify the effectiveness of the algorithm in this paper,and to provide a high-quality marine life image dataset for subsequent marine life target detection.In the marine life target detection part,a lightweight S＿CYOLOv5s network based on the YOLOv5 s network framework,Shufflenet lightweight network and embedded CBAM attention module is designed based on the attention mechanism,5350 marine life images are randomly selected as the training set to train different networks in terms of average precision mean,precision rate,recall rate,After 600 training sessions,the average precision mean value of S＿CYOLOv5s network reached79.99% at a cross-merge ratio of 0.95,which was slightly lower than that of the original YOLOv5 s network by about 0.83%,but the model size was reduced by 14.45 MB,or about 53.9%.The network not only can significantly reduce the model size,but also can ensure high recognition accuracy,achieving a balance of accuracy and speed.