Abstract: Aiming at the problems of target omission in dense scenes and insufficient generalization ability in complex backgrounds for the detection transformer (DETR) series models, we proposed an attention-enhanced Salience DETR model. By embedding a foreground attention mechanism as well as spatial and channel joint attention mechanisms into the Salience DETR model, the model’s ability to extract features of small targets and recognize multiple targets in complex environments was improved. The experimental results on the Common Objects in Context dataset, Vision Meets Drone dataset, and Insects dataset show that the mean average precision of this model is improved  by 1.0 percentage point, 1.2 percentage points, and 2.2 percentage points respectively compared to the Salience Detection Transformer model. The experimental results on a self-collected corn borer dataset show that the model improves  the evaluation metrics for small target recognition  by 2.0 percentage points, verifying the model’s generalization capability. Therefore,  this model enhances detection performance in complex environments and provides a feasible technical solution for small target recognition and specific domain applications.

Key words: small object detection, attention mechanism, complex background, tobacco pests identification