CGW-DETR: a multi-scale feature enhanced RT-DETR Qianxi cherry tomato maturity detection model for dense small targets

Detecting the ripeness of QianXi cherry Tomatoes is critical to intelligent harvesting. There are three major challenges associated with this task: (1) Complex environmental interference makes it difficult to extract surface features from the tomatoes; (2) High rates of missed detections and false positives when detecting densely clustered small targets; (3) High computational costs due to the large number of model parameters and the significant computational load. To address these challenges, this study proposes a lightweight detection model, CGW-DETR, based on an improved RT-DETR architecture. First, we propose a lightweight feature enhancement module, CAS, to optimize target feature perception and suppress interference from complex backgrounds. Second, we introduce a gradient-space pyramid pooling layer (GSPP) that balances local and global feature extraction through multi-scale spatial pyramid pooling and an efficient layer aggregation network, improving the model’s detection of small and densely clustered targets. Finally, we introduce the WIoU loss function and optimize the gradient gain allocation strategy to improve detection accuracy while maintaining the model’s low computational complexity. Experimental results show that, on our self-built QianXi cherry Tomato ripeness detection dataset, compared to the baseline model RT-DETR, the proposed model achieves a detection accuracy of 94.2% (an increase of 2.0%), a recall rate of 88.2% (an increase of 4.4%), an mAP50 of 91.3% (an increase of 3.1%), and an mAP50:95 of 74.6% (an increase of 5.5%). Under the experimental settings of this study, CGW-DETR showed relatively favorable detection results for densely distributed small-scale QianXi cherry tomatoes in complex environments. Its performance was compared with several mainstream models on the self-built dataset.