Research On Methods For Fish Identification Under The Complex Scene Based On Transfer Learning

The overfishing of marine fisheries seriously threatens the ecological security of fisheries.Fishing supervision is one of the main ways to maintain the ecology of marine fisheries.However,the supervisory departments are also facing the predicament that the monitoring means and methods are relatively backward.The intelligent upgrade of the fishing supervision system is an effective approach to solve the problem.Fish identification methods with higher accuracy and generalization are the key to make the supervision system more intelligent.This research mainly used image classification and target detection technology based on transfer learning to explore fish identification methods specifically for complex scene of marine fishery fishing supervision.According to the actual work situation of fishing supervision,the research work was carried out from the scene after fishing vessel landing and the real-time detection scene.Firstly,we solved the problem that the data amount of the image classification task was small.The training data set comes from Kaggle,which is a data science competition platform.In order to make the data distribution more reasonable,the data set was divided into three parts by the method of stratified sampling: training set,verification set and test set.Meanwhile,the data were augmented to avoid overfitting.Because the Lucid Data method required fine data annotation and computing equipment that exceeded actual conditions,we only chose the common methods for data augmentation,including inversion,cropping,and translation.In addition,we used the LabelImg software to label the data for target detection tasks.Secondly,we discussed a fish identification method based on transfer learning and model fusion.A new network to be trained was set up based on the three models of AlexNet,InceptionV3,and ResNet50.The feature extraction parts of the three models was retained,and then new structural modules was connected to them.Pre-trained parameters and un-pre-trained parameters were used to train the network to be trained on the dataset before the data augmentation and the dataset after the data augmentation.In order to further strengthen the robustness of the method,the simple average method in ensemble learning was used to fuse the three models.It is verified by experiments that the improved network trained by transfer learning can raise the identification accuracy by more than 20%,and the improved model is more robust after model fusion.Thirdly,the real-time fish detection method based on YOLO-V3-Tiny-MobileNet was discussed.Compared with the monitoring method for detection after fishing vessel landing,the monitoring method of real-time detection can send out the operation warning to the on-site fishing operators in time,which is more conducive to the protection of fishery ecology.In this study,YOLO-V3-Tiny was used as a baseline to explore a method for real-time fish detection.YOLOV3-Tiny mainly extracts shallow features,and its capability of feature extraction is also insufficient.These problems were improved through MobileNet to form the YOLO-V3-TinyMobileNet network structure.The improved network was pre-trained by VOC2012 and then was retrained on the NCFM dataset.The model obtained was compared with the unimproved network in many aspects.The experiment results show that the average accuracy of the improved model can increase by 5.51% and the parameters can reduce by more than five million,and also show the convergence of the network trained by the pre-trained parameters is 37 epoches ahead of time.The research shows that,compared with traditional methods,the fish identification and detection methods based on transfer learning can more effectively solve the problem of fish identification in complex scene of marine fishery fishing supervision,and can provide reliable technology support for the intelligence upgrading of the fishery fishing supervision system.