Research on Ceramic Surface Micro Defect Detection Algorithm Based on RECS-DETR

In the task of surface defect detection for ceramic materials, identifying minute defects has always posed a challenge due to the difficulty of balancing detection accuracy with computational cost. To address this issue, this paper proposes an enhanced algorithm based on RT-DETR, named RECS-DETR (RDNet-EA Attention-CARAFE-SCFF-RTDETR). First, a lightweight RD feature extraction network is innovatively designed to replace the backbone network of RT-DETR, and Efficient Additive Attention is introduced to substitute the self-attention mechanism in AIFI, thereby reducing computational costs. Next, the CARAFA upsampling module is employed to replace the Upsample module in RT-DETR, which better aggregates information and captures features. Subsequently, a novel structure called SCFF is designed specifically for tiny object detection. The SCFF small object detection optimization module introduces the S2 layer as the small object detection layer and constructs a new feature fusion pyramid, further enhancing the model's sensitivity to minute defects. Finally, a novel composite loss optimization strategy, NWD-Inner-Wise-IoU, is implemented, significantly accelerating model convergence and enhancing the detection of small objects. Experimental results demonstrate that, compared to the ResNet18-based RT-DETR algorithm, the proposed RECS-DETR achieves a 4.4 percentage point improvement in mean average precision on a ceramic surface dataset characterized by tiny defects with black spots. Additionally, it reduces the number of parameters by 45%, decreases FLOPs by 35.8%, and increases detection speed by 17%. The significantly improved performance effectively meets the requirements for detecting minute defects on ceramic surfaces in industrial production.