Seafood Target Detection Algorithm Based On Deep Neural Network

In recent years,the consumption of land resources has been serious,and many countries in the world have listed deep-sea resource exploration as a national engineering project.China has successively proposed national marine engineering and the construction of the 21 st Century Maritime Silk Road,aiming to strengthen the country’s ability to control resources within the territorial waters and serve the national economy.Underwater target detection is the key technology to realize the intelligent exploitation of marine resources,which is related to the control of underwater robots and the efficiency of resource exploitation.Therefore,underwater target detection has become a research hotspot in the field of computer vision.Underwater target imaging is usually accompanied by color deviation,accumulation occlusion,high similarity between foreground and background,and multi-scale problems.Therefore,the general target detection algorithm is directly used in underwater scenes,often can not achieve ideal results.In this paper,the target detection algorithm of multi-class seafood optical image in underwater real scene is studied in depth from the characteristics of data set and algorithm optimization.The main work of this paper is as follows:(1)With the data set provided by the 2020 National Underwater Object Detection Algorithm Contest as the original data,a variety of underwater image enhancement technologies are used to complete the clarity processing.Afterwards,non-target samples and interference samples were removed,the number of various seafood samples in the dataset was counted,stitching and screenshots were used to perform undersampling and oversampling,to make the distribution of samples in the dataset more balanced.Finally,label the preprocessed data set to obtain the corresponding labeling file,divide the data set according to the proportion to obtain the training set and test set,determine the baseline algorithm Faster R-CNN used in this research through experiments.(2)For the problem of underwater occlusion target detection,based on the triplet attention mechanism,a dilated convolution module is proposed to replace the original Z pooling module to reduce the loss of fine-grained information.Based on the non-local neural network,the Gaussian similarity function and the concatenation similarity function with the visual reasoning function are weighted integrated,to endow the non-local neural network with certain logical reasoning ability.Finally,the spatial attention branch in the triplet attention is replaced by the improved non-local neural network to form a new triplet attention,to alleviate the problem of detection accuracy degradation caused by occlusion.(3)To address the problem of underwater multi-scale target detection,this paper focuses on the analysis of the structure of feature pyramid network.A new feature pyramid network,named coordinate trident feature pyramid network is proposed.The trident feature enhancement module and the coordinate non-local module proposed in this paper,are inserted into the horizontal connection and upsampling process of the original structure,to further enhance the representation ability of the feature pyramid for multi-scale features.(4)For the problem of inaccurate positioning of seafood targets,the linear regression loss gain coefficient is introduced in the regression loss function L1-smooth of Faster R-CNN,which is guided by the intersection over union.The purpose of this coefficient is to increase the penalty of the offset between the prediction box and the real box,accelerate the regression speed of the prediction box,and make the position regression of multi-scale targets more accurate.Linear form is beneficial to training convergence.(5)The effectiveness of the proposed algorithm is proved by comparing with other algorithms for underwater seafood target detection.This paper contains 62 figures,17 tables and 81 references.