Abstract: Calculating the Rock Quality Index (RQD) based on the results of geological core sampling to evaluate the distribution of faults, discontinuities, and fractures in the dam site area is of great significance for quality management in water conservancy project construction. However, manual measurement of RQD is time-consuming, labor-intensive, and minimally automated. Additionally, the accuracy and practical effectiveness of image-based RQD recognition technology are limited due to complex interferences such as lighting and shadow effects, texture color differences, and manual markings, as well as the challenges in identifying the effective length of the core caused by small cracks. To address these issues, this paper conducts a study on improved RQD automatic identification using YOLOv8-seg with enhanced small crack recognition under complex interference. This model builds on the YOLOv8-seg framework and introduces a lightweight multi-semantic space and channel collaborative attention module, SCSA, to accurately extract core features despite the interference from shooting conditions and manual markings, while incorporating dynamic serpentine convolution to enhance the model's ability to detect small cracks, thereby achieving effective extraction of core sections. Furthermore, based on the image segmentation results of core sections, RQD is automatically calculated and recorded. Case analysis reveals that, compared to the original model, the improved model increases precision (P), recall (R), mAP50, and mAP50:95 by 3.3%, 2.2%, 2.4%, and 1.9% respectively. The proposed method is more efficient than manual calculations, achieving an average relative error of 1.21%, thus providing a more robust and reliable technical scheme for RQD recognition calculation and rock mass quality evaluation.