Research On Small Target Detection Algorithm For Fish Shoal Based On YOLOv5

Small target detection of fish schools is one of the categories of target detection.Target detection has also made great progress in the context of deep learning and the wide application of neural networks.Underwater target detection based on deep learning and neural network has also attracted the attention of scholars around the world.However,the imbalance in the type,shape,quantity and size of underwater targets will lead to poor robustness and low accuracy of traditional target detection algorithms.In order to overcome the problem of false detection and missed detection of small fish in complex underwater environment by traditional algorithms,many scholars have proposed underwater target detection algorithm based on depth learning in recent years.In the task of underwater small target fish detection and recognition,scholars at home and abroad have proposed many methods,but there are still three problems: first,the existing model is relatively bulky,with many parameters,low detection speed,and the accuracy needs to be improved.Second,the existing models do not make full use of the high and low level features of the image,and feature extraction is not comprehensive enough.Third,the detection accuracy of existing models in natural scenes is low and is greatly affected by the environment.Aiming at the above pain points,an improved YOLOv5 target detection method(INV-YOLOv5)is proposed.The specific research contents and main problems to be solved include the following parts:(1)A dataset has been established for detecting small targets in underwater fish schools.(2)Replace the Focus module in YOLOv5 m with a convolutional module;And added inner convolution operators to the C3 module of the original backbone network(Backbone);Introducing multi head sub attention modules into the backbone network;The experiment proves that the improved algorithm has a certain improvement in the detection and recognition rate of fish school targets.(3)In order to better approach the commercialization implementation,this article uses PyQT5 to build a YOLOv5 user interactive visual interface based on the proposed algorithm.In the experimental stage,the precision,recall and mean average precision(mAP)were selected to evaluate the model.The experimental verification on the Labeled Fish in the wild dataset and Wild Fish dataset shows that the average accuracy(mAP)of the improved YOLOV5 method(INV-YOLOV5)is 81.7% and 83.6% respectively,which is improved by 6 percentage points and 14.5 percentage points compared with YOLOV5 network.It not only has a higher recognition rate but also is more lightweight.The model size is reduced by about 6M(Mega)compared with YOLOV5 network,which meets the requirements for small target detection in complex environments.