| The total amount of marine fishery resources worldwide has been decreasing year after year,and the protection and management of marine fish have attracted much attention.In order to develop marine fishery resources reasonably and sustainably,it is of great significance to design an efficient and accurate fish detection and recognition system.Traditional fish detection and recognition methods mainly rely on the features of artificial design,but these features are usually not universal,and the design of artificial features requires expert-level experience.To solve the above problems,we propose a detection and recognition scheme based on deep learning for fish recognition task in the scene of ship borne electronic monitoring,which mainly includes the following two aspects:Firstly,we propose a real-time fish detection case based on multi-scale feature fusion.The contents of specific work can be summarized as follows:first,we design a multi-scale feature fusion module based on SSD detection algorithm,which can effectively fuse the features of deep layer containing context high-level semantic information and the high-resolution features of shallow layer containing more detailed information of the target.Constructing a multi-scale feature pyramid inside the network effectively improves the fish target detection effect(especially the improvement of small targets is particularly obvious).Secondly,different scale and number of default boxes are designed on each feature map of the prediction layer,which can reasonably allocate the computing resources and solve the detection problem caused by small fish target.Finally,for the problem of high false detection rate and poor detection accuracy of difficult samples,we combine the advantages of cross entropy loss function and focus loss function to make the difficult and easy samples fully trained in the training process,so that the learned model is more robust.Experiments show that the proposed fish detection algorithm based on multi-scale feature fusion has achieved good results in both accuracy and efficiency.The map is83.08%,and the processing efficiency is 24.54 ms per sample.Secondly,we propose a two-stage fish recognition scheme to further improve the accuracy of fish recognition.The solution consists of two parts:region proposal module and image recognition module.For an image,region proposal module gives target proposal regions, and then image recognition module classifies all proposal regions.The region proposal module mainly refers to the RPN module in Faster R-CNN algorithm.In the design of image recognition module,a special Inception module is designed for fish recognition problem,and then the image recognition network(Inception-ResNet)was designed by combining the Inception module and the ResNet.Experiments show that the two-stage fish recognition scheme designed in this paper can improve the accuracy of fish recognition,and the top-1 is 91.48%.In conclusion,the fish detection scheme based on multi-scale feature fusion proposed in this paper can achieve better detection effect and satisfy real-time sample processing.The two-stage fish recognition scheme can further improve the fish recognition accuracy. |
