<p>Facing the global public health challenge of the high prevalence of chronic diseases and an increasing healthcare burden, food image-based dish detection plays a vital role in personalized nutrition regulation and intelligent dietary management. To address the challenges in Chinese dish detection, including complex backgrounds, multiple coexisting dishes, difficulty in detecting small objects, and poor adaptability of existing methods, this study proposes a task-oriented detection network named CRD-DETR. The proposed method integrates three complementary modules—Coordinate Attention (CA), Receptive Field Attention Convolution (RFAConv), and Dynamic Position Bias (DPB)—at different representation stages to enhance local feature extraction and spatial relationship modeling. To validate the effectiveness of the proposed method, experiments were conducted on a self-constructed and manually annotated dataset named CDish39, which was collected in a real-world cafeteria environment. The results demonstrate that CRD-DETR achieves a 4.7% improvement in the mAP@0.5:0.95 metric compared to baseline and state-of-the-art models. Furthermore, ablation studies confirm the individual and combined effectiveness of the integrated modules. CRD-DETR achieves accurate detection of dishes in both single- and multi-dish scenarios, significantly reducing missed detections, false positives, and duplicate predictions.</p> Graphical Abstract <p></p>