Abstract: In response to the problems of low detection precision and poor real-time performance caused by the extreme similarity between object and background in complex scenarios, an object detection algorithm based on improved RT-DETR is proposed. The influence of FLOPs (floating point operations) and FLOPS (floating point operations per second) on real-time performance is analyzed. FasterNet module, which is excellent in speed, efficiency and lightweight, is used as the backbone network to realize a lightweight network structure and real-time algorithm improvement. The HiLo (high- and low-frequency attention) module, which can process high-frequency and low-frequency image features simultaneously, is introduced as the intra-scale fusion module to improve algorithm detection precision and anti-interference ability. Experiments are carried out on CAMO-D, CottonInsect and a self-built pest and disease dataset in complex scenarios, and the results show that the improved algorithm is superior to the RT-DETR (real time detection transformer) algorithm in that the detection precision has increased by 8.0%, 6.6%, and 9.0%, and F1 score has increased by 13.8%, 4.2%, and 3.2%, respectively. Under the premise of frame rate being unchanged, the floating-point operations of the improved algorithm have reduced by 49.6%, and number of parameters reduced by 45.7%, which makes the calculation accuracy and efficiency improve significantly. The algorithm effectively addresses the challenge of object detection under similar background interference and holds promise for broader application in various practical scenarios.