An accurate inland water garbage recognition network for USV camera images

Abstract Accurate detection of water surface garbage is crucial for developing an environmentally friendly Internet of Things (IoT) system based on unmanned surface vehicles (USVs). However, it is still challenging to automatically recognize and measure the location of water garbage, hindered by complex factors like varying sunlight conditions and the minute size of garbage targets. This paper aims to develop an accurate water garbage recognition network (WGR-Net) that improves performance through efficient feature extraction, transmission, and restoration of feature resolution. The proposed method first adopts the YOLOv9 network architecture that combines generalized efficient layer aggregation network with programmable gradient information to overcome the problem of data loss in deep networks. Then, in order to improve the accuracy and training efficiency of models with massive parameters, the backbone module of the pretrained model on the COCO dataset is frozen for feature extraction. The head module of this pretrained model is transferred and fine-tuned by USV camera images specifically for water surface garbage recognition. Furthermore, an ultra-lightweight and effective upsampler is introduced into the fine-tuned model to restore the feature resolution. The performance of the proposed model is tested using the FLoW-IMG dataset collected by the ORCA unmanned cleaning vessel and WSODD dataset, and comprehensive performance comparisons are conducted on multiple YOLO series models. The results demonstrate that the proposed WGR-Net significantly improves the accuracy of water garbage recognition, achieving a mAP@0.5 of 92.9% and mAP@0.50.95 of 51.7%. The garbage tracking results of water surface video also show a reduction in missed and false detections. The proposed method effectively promotes the accurate recognition of inland water garbage, providing strong technical support for the application of USV based environmental IoT systems.