Research On Classification And Detection Of Underwater Organisms Based On Deep Learning

The classification and identification of marine organisms such as fish,sea urchins,and corals is conducive to the management of marine biological systems and biodiversity,as well as the analysis of species differences in marine organisms and the protection of endangered marine organisms.Studying the distribution of various marine organisms helps to analyze the impact of global warming and human exploitation of marine resources on marine life,thereby guiding human rational exploitation of marine resources.However,due to factors such as sediment,suspended particles,light absorption and scattering,the image presentation of underwater biological datasets is poor,with color distortion,and reduced contrast,greatly affecting the identification characteristics of targets.Secondly,the proportion of complex underwater environment background in the image far exceeds the area of aquatic biological targets,further affecting the difficulty of locating the target location.In order to deal with the complex underwater environment of target recognition,we conducted experimental research.This includes filtering noise from underwater images rich in noise or with complex backgrounds during the feature extraction stage of the image and improving the accuracy of target classification during the recognition process.The following two aspects of work have been carried out:(1)An underwater target classification algorithm(FCMFDA-ELM)based on an improved flow direction algorithm and a search agent strategy that can simultaneously optimize the weight parameters,bias parameters,and hyperparameters of an extreme learning machine(ELM)is proposed as a new underwater target classifier to replace the fully connected layer in traditional classification networks to build a classification network.In the first stage of the network,the Dense Net-201 network trained by Image Net is used to extract features and reduce dimensions of underwater images.In the second stage,the optimized ELM classifier is trained and predicted.In order to weaken the uncertainty caused by the random input weights and biases of the introduced ELM,a chaos based initialization,multi group strategy,and fuzzy logic optimized flow direction algorithm(FCMFDA)are used to adjust the input weights and biases of the ELM while optimizing its hyperparameters using a search agent strategy.Through comparative experiments on Fish4 Knowledge and Underwater Robot Professional Competition(URPC)datasets,it can be seen that the FCMFDAELM classifier proposed in this paper effectively improves the classification effect for complex underwater images.(2)A residual building unit(RBU-TA)with two-terminal attention mechanism is proposed.By introducing a reinforcement channel attention mechanism into the direct mapping edge of the residual structure,adaptive compression of noise rich feature map channels can provide rich shallow image information for high-level deep convolution features while avoiding shallow noise pollution.In addition,to address the problem of class imbalance in underwater biological images,a difficult sample resampling combined with a Focal Loss function is used to suppress an excessive number of negative background samples,and to retrain proposed targets that are difficult to distinguish.Improve the detection accuracy of the network for rare and indistinguishable underwater organisms.The method was validated on the URPC and RUOD datasets,and the results showed an average detection accuracy of 61.26% and72.81%,achieving stable accuracy improvement.Effectively improve target detection for organisms in complex underwater environments.