Abstract: To address the challenge of low detection accuracy caused by the small size, curled shape, and complex natural background of insect-bitten Zijin Cicada Tea, an improved target detection model based on RTDETR-R18, named RTDETR-IM, was proposed. In this model, the original backbone network was replaced with CSPNet, and a newly designed FGMA module was introduced to replace the C2f residual structure. Adaptive feature fusion based on the Fourier transform and the AFG mechanism was incorporated to significantly enhance the ability to detect small targets. Additionally, the original feature fusion module was substituted with a MSPCA module, and a context-guided feature reconstruction strategy was applied to improve detection performance under cluttered backgrounds. Experimental results showed that, compared to the baseline model, an increase of 5.2% in overall detection accuracy and an 8.4% improvement in mAP_0.5 were achieved. At the same time, the parameter count and computational load were reduced by 3.62 M and 7.7 GFLOPs, respectively. Through this approach, both high detection accuracy and enhanced computational efficiency were ensured, making the model well-suited for real-time detection of insect-bitten Zijin Cicada Tea in practical agricultural scenarios. The findings provide theoretical support for the detection of insect-bitten Zijin Cicada tea buds in agricultural scenarios. Key words: Small target detection，RTDETR，Feature fusion，Zijin Cicada Tea