Distant Water Surface Garbage Recognition Method in Complex Scenarios

To improve the insufficient identification rate of remote garbage targets by water surface cleaning equipment, a model based on RT-DETR for remote garbage target detection called FRT-DETR was developed. First, CSPCAA-ADown was proposed as a modified backbone network, which better captures feature information of garbage targets beyond 30 meters at different scales while reducing computational burden, thus minimizing the impact of insufficient feature information caused by distance. Second, a CCGAF module was introduced to dynamically adjust the importance of features from different layers. Finally, an adaptive wavelet pooling module, WaveletPool, was employed for upsampling and downsampling feature maps, thereby reducing the impact of target scale variations and achieving good recognition rates even under conditions of blurry distant targets and complex backgrounds. The improved model achieves an mAP0.5 of 91.3% on the FWSG dataset with a reduced parameter count of 11.8M, providing a high-precision, high-efficiency, and low-computational-cost detection solution for water environment monitoring.