Research On Fish Fine-grained Image Recognition Algorithm Based On Deep Learning

With the changes in the global marine environment,fish and fish habitats are facing increasing pressure.In terms of fish farming,due to the threat of alien species,the production of fish farming will decrease.The detection and identification of underwater fishes for better species classification becomes crucial.Relying only on artificial classification of underwater fish is often destructive and time-consuming.It can help scientists obtain the distribution of certain specific fish through automatic fish identification,monitoring and evaluation of fish populations,and analysis of changes in the marine environment..In the complex underwater environment,the images of many fishes are not clear,the features are not obvious,and the number of effective images acquired underwater is small and the amount of data is uneven.In addition,many fishes cause the same due to different illumination,shooting angles and postures.There are big differences within the species,but small differences between different species.Therefore,the research on the detection and recognition of fine-grained images of fish has important theoretical significance and practical value.This paper studies the detection and recognition technology of fine-grained images of underwater fish.The main research is as follows:(1)In order to facilitate more effective detection and recognition of fine-grained images of underwater fishes,a priori defogging algorithm based on dark channels is used to preprocess the data set of underwater fishes.The images taken underwater are prone to blurring due to the scattering of light by impurities in the water,and this phenomenon is the same as the generation of haze in the terrestrial environment,so the dark channel prior defogging algorithm is selected to deal with underwater The image is blurred.In this paper,the pre-processed data set based on the dark channel prior defogging algorithm is used to study the detection and recognition of underwater fish fine-grained images.(2)In order to quickly locate and identify fishes photographed underwater,the Faster R-CNN algorithm fused with feature pyramids is designed to detect underwater fishes.The algorithm first uses Resnet101 as a feature extraction network,and then fuses the feature pyramid network with Faster R-CNN,and predicts by fusing the information with higher resolution in the low-level feature map and the information with richer semantics in the high-level feature map.The NMS algorithm in the Faster R-CNN algorithm is improved to the Soft-NMS algorithm,which reduces the confidence of the candidate frame by linear weighting so as to ensure that fish with serious underwater overlap will not be missed.This paper verifies through experiments that the average accuracy of the fish detection algorithm has reached 91.6%.(3)In order to solve the problem that the fine-grained characteristics of underwater fish are not obvious and the sample data of fish is not balanced,the FL-BCNN fine-grained fish recognition model based on feature fusion is designed.The model uses VGG16 for feature extraction and improves the structure of the bilinear convolutional neural network to better obtain the fine-grained features of underwater fish.Finally,the focus loss function is used to replace the original standard cross-entropy loss function.Solve the problem that the model is prone to overfitting due to the unbalanced sample data.This paper verifies through experiments that the algorithm model has an accuracy of 97.68% for the recognition of fine-grained images of fish.(4)Using the improved Faster R-CNN fish detection model and FL-BCNN fish fine-grained image recognition model,a fish fine-grained image detection and recognition software was developed based on the Python language,which realized underwater fish The visualization and result storage of class fine-grained image detection and recognition lays the foundation for the subsequent statistics and analysis of fish species.