Research On Detection Method Of Damage Aquaculture Net Based On Deep Learning

In aquaculture,net can effectively control the range of activity of fish groups and facilitate centralized management.However,the structure of net is complex,and it is highly susceptible to damage from natural weather such as wind and waves.If not detected and repaired in time,a large number of fish groups could escape,causing significant economic losses.Traditional net damage detection requires professional divers to conduct underwater surveys,which poses issues such as high cost,time consumption,and personnel safety.This article used digital image processing technology and deep learning technology to study the recognition method of the location and degree of underwater net damage.The research results were as follows:(1)Establishment and enhancement of a damaged net dataset.The fifish v6 ROV equipped with the camera captured images and videos of the net from multiple angles and distances in daytime and nighttime environments.By extracting the key frames of the captured video and the captured images,we obtained different resolutions of net images.The obtained images are manually labeled with different labels based on the location and degree of damage,and two sets of datasets were established.Enhanced the established dataset through digital image processing technology to enhance image contrast and expand the dataset.(2)Deep learning-based detection of damaged areas in net.Based on one-stage detection models YOLOv5 s,YOLOv5m,and two-stage detection model Faster R-CNN,training was conducted on the dataset of damaged areas in net.The results showed that YOLOv5 m had the best evaluation indicators and its accuracy,recall,m AP,and average recognition rate were 95.29%,96.06%,98.47%,and 11.34 frames/s;The evaluation indicators of YOLOv5 s took second place,whose accuracy,recall,m AP,and average recognition rate were 93.67%,93.64%,98.05%,and 34.02 frames/s;The evaluation indicators of Faster R-CNN were the worst,whose accuracy,recall,m AP,and average recognition rate of 72.38%,98.57%,95.56%,and 0.34 frames/s,which was significantly different from the previous two models.Due to the slow recognition time of YOLOv5 m and its difficulty in real-time detection,the YOLOv5 s training strategy had been improved to achieve better results.By transforming the size of the training set images and using methods such as cosine-warmup learning rate attenuation,the improved YOLOv5 s training strategy improved accuracy,recall,and m AP by 1.52%,2.42%,and 0.46%,its average recognition rate was 34.10 frames/s.(3)Deep learning-based detection of damaged degree in net.Based on different damage situations,dividing the degree of damage net into four target categories: edge damage(Broken),two holes(Broken＿2),three holes(Broken＿3),and more than three holes(Broken＿large).YOLOv5 s,which has been improved through the training strategy,was used for training.The accuracy of edge damage,two holes,three holes,and more than three holes after training were 73.48%,64.71%,91.59%,and 98.99%,The recall were 97.98%,99.00%,98.99%,and 98.99%,respectively.After training with Faster R-CNN,the accuracy for edge damage,two holes,three holes,and more were 58.33%,61.54%,91.30%,and67.44% The recall were 36.48%,99.12%,84.00%,and 96.67% The evaluation indicators were significantly weaker than YOLOv5 s.In order to optimize the detection performance of the model,the YOLOv5 s model was improved by adding an attention mechanism to the original network structure.After testing,adding CBAM attention mechanism to the 8th layer of the model’s backbone increased the recall of model edge damage by 1.02%,the accuracy of breaking two holes increased by 3.10%,the accuracy of breaking three holes increased by 0.99%,the recall increased by 0.14%,and the average recognition time was36.23 frames/s.(4)Development of a damaged net detection software platform.A damaged net detection software platform was designed based on Qt Designer and PyQt5.The software can use deep learning models to detect the location and degree of net damage in images or videos,and achieve different detection effects by modifying detection parameters.The test results showed that the software interface can meet the detection requirements and had certain practical value.