EAS-DETR: an improved RT-DETR for object detection in quadruped robot inspection images

Abstract During inspection, onboard vision systems of quadruped robots are frequently affected by viewpoint jitter, local motion blur, background interference, and substantial object-scale variation. These factors degrade object boundaries and local textures, interfere with contextual feature representation, and increase the difficulty of multi-scale feature modeling. To address these challenges, this paper develops EAS-DETR based on RT-DETR by improving backbone detail extraction, low-redundancy intra-scale interaction, and cross-scale feature fusion. First, an Efficient Detail Enhancement Module is introduced to strengthen local structural and texture responses degraded by platform motion and local blur. Second, a partial-channel singlehead self-attention module is introduced into the intra-scale interaction stage. Self-attention is applied only to a subset of channels, while the remaining channels preserve local responses through a bypass branch, thereby reducing encoder computation and memory-access overhead. Finally, a redesigned multi-scale feature pyramid network is introduced to reduce representation mismatch during hierarchical feature fusion and improve the representation of objects at different scales. Experimental results on the UPTD dataset showed that, compared with RT-DETR, EAS-DETR reduced the number of parameters and FLOPs by 37.1% and 24.6%, respectively, and increased the inference speed from 237.75 to 266.29 FPS. Precision, recall, F1-score, mAP 50 , and mAP 50:95 increased by 1.28, 2.36, 1.81, 2.67, and 1.69 percentage points, respectively. These results demonstrated that EAS-DETR achieved a more favorable balance among detection accuracy, model complexity, and inference speed in complex quadruped robot inspection scenes.