| The 21st century is considered to be a period of large-scale development and.utilization of the ocean.One of the ways for China to build a marine power is to develop the marine economy and improve the ability to develop and utilize marine resources.As a pillar industry of marine economy,fishery has developed rapidly in recent years.Facing the complexity of the marine environment,due to the limitations of human physiological characteristics,strong equipment and technology are needed to replace them to complete high-intensity underwater operations.A variety of underwater robots can flexibly design functions according to needs and provide help for the field of aquaculture.However,how to quickly and accurately identify different seafood is an urgent problem to be solved.For small objects in close range,optical image detection shows its unique advantages.In recent years,image detection based on deep learning convolutional neural network has made great progress and showed excellent performance.However,due to the limitation of underwater visibility conditions and the interference of complex marine environment such as the scattering and absorption effect of water on light,the brightness of underwater image will decline,and there will be problems such as fuzzy characteristics,low contrast and distortion.At the same time,it is difficult to collect underwater image and lack of sufficient training data,which increases the difficulty of visual detection.Therefore,when moving the trained target detection model to underwater target detection,we need to solve the problems of insufficient data sets and low underwater image quality.Aiming at the problem of uneven distribution of sample categories in the data set,for the classes with a small number of samples,the underwater style image containing targets is generated by neural style transfer network,and the similarity between the underwater image generated by neural style transfer network and the real image is evaluated by maximum mean difference and similarity measurement.Different processing methods are adopted for the generated images in different situations.The generated images with the target occupying most of the image area are fused into the background of the underwater image through Poisson image editing.The final generated image is labeled to realize the expansion of the data set.Aiming at the problem of low image quality when convolutional neural network is used to detect underwater targets,an improved yolov4 network model is used to improve the effect of underwater target detection.Firstly,the characteristics of underwater image data and the structure of yolov4 network are analyzed.Aiming at the problems of image blur and small samples,the output structure of the network is improved based on yolov4 model,a larger scale output is added,and the convolution attention module is introduced to enhance the effective features and reduce the interference of background information.The experimental results evaluate the performance of the improved algorithm based on yolov4 in underwater multi-target detection,and verify the effectiveness of the improved network based on yolov4. |
