Research On Ocean Fish Detection Technology Based On Deep Learning

Marine fish detection is one of pivotal technologies for exploring and utilizing marine resources.Accordingly,studies on accurate and fast detection for marine fish have attracted much attention and become a subject of great significance with regard to marine resources protection and sustainable development.Traditional fish detection mainly relies on artificial design features and machine learning classification.However,due to the complexity and variability of underwater environment,traditional detection algorithms share several common limitations,such as the difficulty of artificial feature design and poor portability and robustness,which hardly fulfil realistic requirements.With the rapid development of computer science,deep learning is widely used in detection fields,such as trajectory tracking and biological detection,greatly improving the detection accuracy in related fields.In this manuscript,aiming at the disadvantages of traditional fish detection,deep learning-based target detection is systematically studied and employed to detect marine fish.The main research contents and conclusions are summarized as follows:1.The disadvantages of the SSD network are analyzed.Moreover,a multi-scale feature-enhancing SSD network(FE SSD)is proposed to address these issues,aiming at the defects that shallow feature layers in the SSD network have small receptive field and lack semantic information and contextual connections,resulting in the poor effectiveness of multi-scale target detection.The FE SSD network improves the multi-scale information of the network and enhances the network context by using the FPN structure and feature enhancement module,thus improving the detection accuracy of multi-scale targets of network.In order to fulfil the real-time requirements of marine fish detection,several techniques,such as volume integral solution and grouped convolution,are employed in FE SSD to reduce network parameters as much as possible and accelerate the detection speed.2.In this study,we analyze the influence of uneven distribution of database for the detection accuracy,containing the uneven distribution of category data and the uneven ratio between hard-to-separate samples and easy-to-separate samples.Aiming at the issue of uneven distribution of category data,several strategies of oversampling and equalizing the data set are employed to increase the contribution of the minority category during model training and improve the detection accuracy of the FE SSD network for the minority category.Regarding the issues that the number of difficult-to-separate samples is usually little,but they are more significant for training model,Focal loss is adopted to focus on these difficult samples to improve the effectiveness of the model in terms of target detection.To well integrate Focal loss into the FE SSD network,plenty of experiments are implemented to determine the optimal parameters of Focal loss for the proposed FE SSD network.Finally,comprehensive experiments validate that the data equalization method can effectively alleviate the negative effect of uneven data,and improve the detection accuracy of the FE SSD network for marine fish.3.In order to meet the requirements of realistic detection,a marine fish detection system based on MFC is constructed,which specifically includes image acquisition,image detection and effect exhibition modules.The marine fish detection system is characterized by convenient operation,and is capable improving the automation process of marine fish detection.