Research And Application Of Fish Detection Based On Deep Learning

China has a vast marine territory and has extremely rich marine resources.The study of fish detection and identification can better assist fishing and water quality monitoring,and provide theoretical basis for fishery resources protection and management.In recent years,the research in the field of underwater video images has been continuously deepened,which provides a reference for the detection and recognition of fish.However,the complex background of underwater video image and the endless variety of fish forms make the fast and accurate positioning and identification of fish targets face many challenges in underwater environment.Most of the existing detection and identification methods are based on supervised learning,which requires a lot of manpower and material resources.With the development of artificial intelligence,deep learning has become a hot topic in the field of image recognition with its powerful internal network automatic feature extraction capability and highprecision recognition effect,which provides new ideas and new methods for fish detection and recognition.After studying the related theories of deep learning,this paper constructs a convolutional neural network for underwater fish video images to promote the intelligence of fishing supervision system.The specific content includes the following aspects:(1)In this paper,the freshwater fish dataset Fish30 Image was collected and produced.The dataset included a total of 4737 fish images with complex backgrounds of 30 species such as snapdragon,cartographic fish and yellowfin pompanus.Residual network transfer learning method was used to train Fish30 Image data set and seawater fishdata set Fish4 Knowledge containing a total of 27 370 pieces of 23 species.After that,the final fish species classification results were obtained through softmax classifier.The results of the identification test on 23 fish species showed that t best fish recognition accuracy can be achieved by fixing the parameters of conv1 and conv2 of resnet-50 pre-training model on Imagenet dataset and fine-tuning the high-level parameters.On the public Fish4 Knowledge dataset,the model has achieved the highest recognition accuracy,with an average recognition accuracy of 99.61%.The comparison with other convolutive neural network methods shows that this method has great advantages in recognition accuracy and time performance on both Fish4 Knowledge and Fish30 Image data sets.(2)In order to overcome the problems of insufficient sample size of underwater fish images and the difficulty of detection and low accuracy of small target fish bodies,this paper proposes an improved Retina Net multi-target fish body detection method.Based on deep network transfer learning technology,this method used the combination of deep separable convolution and Retina Net network to learn the multiple scale characteristics of the small fish target,so as to reduce the network parameters and enhance the robustness of detection.By comparing the configuration schemes of the three kinds of deep separable convolution in the Retina Net network,it can be seen that using deep separable convolution to replace standard convolution on the P6 layer of Retina Net network achieves the optimal fish detection effect,with an average detection accuracy of 96.07%.(3)Based on the above research and work,this paper developed an Android-based mobile underwater fish recognition system,and tested the system through actual scenarios.The test results show that the system can realize the accurate identification and classification of underwater fish,and the operation speed is fast and the use is convenient.