Automated detection of hazardous household waste in complex sorting scenarios: A lightweight and high-precision deep learning approach.

Accurate hazardous household waste detection is critical for intelligent municipal solid waste sorting, as missed hazardous items may cause equipment damage, secondary pollution, and safety risks. However, complex backgrounds, object occlusion, scale variation, and high visual similarity between hazardous and ordinary waste make reliable detection challenging. To address these issues, this study proposes CE-DFINE, an enhanced detector based on D-FINE for hazardous waste identification. The model introduces three key improvements: a lightweight feature extraction module to reduce redundant computation, a hypergraph-based high-order semantic association mechanism with multi-scale channel attention to enhance contextual modeling under occlusion and deformation, and a selective attention decoder to suppress cluttered background interference. Experiments on the self-constructed HW-Dataset show that CE-DFINE achieves 86.3% AP50 and 58.8% AP50-95, outperforming the baseline D-FINE by 3.4 and 1.9 percentage points, respectively. Generalization experiments on the external TrashCan dataset demonstrate that the improved model maintains competitive detection performance in more complex environments, such as underwater waste scenes, and shows enhanced capability for small-object detection. Edge deployment on Raspberry Pi 5 achieves 10.40, 5.59, and 3.84 FPS at 320 × 320, 512 × 512, and 640 × 640 input resolutions, respectively. These results indicate that CE-DFINE provides an accurate, robust, and deployable vision-based detection solution for low-cost hazardous waste sorting applications.