EA-DETR: Edge-Aware Detection Transformer for Water Surface Floating Object Identification

Designed to address the challenge of identifying water surface floating objects with transparent features under complex backgrounds such as irritating light sources, we introduce Edge-Aware Detection Transformer(EA-DETR), a novel end-to-end object detection model based on DINO-ResNet50. Three innovative strategies are mainly incorporated. First, Edge-Aware Denoising(EAD) Module, composed of convolutions and Squeeze-and-Excitation(SE) Module, accentuates boundary information across multi-scale features after Backbone of ResNet50. Secondly, Mamba-like Linear Attention(MLLA) Module continuously suppresses incorrect positional weights while emphasizing correct ones through linear operations in Transformer Encoder layers. Finally, a dynamic matching approach combines flexible coefficients for CIoU, a regularization item based on bounding box sizes and a frame utilizing Sober operator to convert coordinated into edges, thus capturing more detailed information of transparent objects by calculating extra segmentation mask loss fostered by boundaries. EA-DETR achieves 72.4% AP(0.5:0.95), 98.7% AP(0.5) and 65.7% AR(0.5:0.95) on the FloW-ImG Dataset, significantly outperforming other DETR-like models. Furthermore, it demonstrates ideal generalization and robustness, with results on the Trash_ICRA19 Dataset exhibiting 44.1% AP(0.5:0.95), 74.2% AP(0.5), and 64.5% AR(0.5:0.95). Overall, the architecture proposed can provide effective tools for learning ambiguous boundary features without disturbance of extreme lightness caused by water surface and the sun.