CHDPL-Net: a lightweight network for Chinese herbal decoction pieces detection

To advance the integration of traditional Chinese medicine (TCM) with next-generation information technologies, the intelligent identification of Chinese herbal decoction pieces (CHDP) has become a crucial research direction. However, the performance of current algorithms remains unsatisfactory. To address this, we have constructed a diverse CHDP dataset and proposed a lightweight network for CHDP detection, named CHDPL-Net. Based on YOLOv8, this model introduces a new network scaling factor to reduce redundant channels in deep feature maps and optimizes the Neck and Head structures to better accommodate CHDP detection, which primarily involves medium and large targets. Additionally, a newly designed downsampling module, RDown, replaces conventional downsampling methods to reduce computational overhead, while the adopted upsampling module, DySample, significantly enhances the recovery of detailed features. To further improve lightweight performance, we apply GhostConv to optimize the SPPF and C2F modules and incorporate a novel attention mechanism, EHA, which makes the model more sensitive to color and texture information, mitigating the performance degradation caused by lightweight design. Ultimately, CHDPL-Net achieved excellent results with only 31.9% of the Parameters and 30.6% of the FLOPs compared to YOLOv8, obtaining $$mAP_{50}$$ and $$mAP_{50:95}$$ scores of 98.2% and 95.4%, respectively, with only a 0.8% performance drop. This demonstrates that the model can meet practical detection needs to a certain extent.