Lightweight Similar Object Detection Method Based on Improved RT-DETR

The lack of feature information in similar object detection tasks generally leads to high false detection rates and large model sizes. To deal with such an issue, in this paper, a lightweight similar detection (LS-DETR) model is proposed to enhance detection accuracy and maintain detection speed. It utilizes the lightweight network architecture generalized efficient layer aggregation network (ELAN) as the main feature extraction backbone network, ensuring high accuracy while maintaining network lightweightness and speed. Furthermore, by combining cross-stage partial (CSP) networks with diverse branch blocks (DBB) and introducing the ELAN, the DBB-CSP-ELAN module is constructed to enhance the network's multi-scale feature extraction capability and reduce model computation. Experimental results on a solid color puzzle dataset demonstrate that, compared to the original real-time detection transformer (RT-DETR) model, LS-DETR achieves an improvement of 4.7% in mAP, a reduction of 53.5% in flops, and a decrease of 53.1% in model size. Compared to other mainstream algorithms, the improved method exhibits significant advantages in terms of model accuracy, size, and computational complexity, meeting the requirements of lightweight similar object detection tasks.