A Lightweight Grayscale-Guided Visual Perception Network for Jadeite Transparency Detection

Jadeite transparency grading is challenging because adjacent categories exhibit high visual similarity and commercial images often contain complex backgrounds and acquisition-related variations. This study proposes a lightweight grayscale-guided visual perception network based on YOLOv10s for joint jadeite localization and transparency-category grading. First, the Jade Texture Layered Convolution (Jade-TLC) module replaces the initial CBS layers and combines asymmetric convolution for slender linear textures with dilated grouped convolution for regional granular features. The two branches are adaptively fused using the whole-image grayscale mean as an auxiliary image-level statistic rather than a physical measurement of transmittance. Second, an improved SimBAM module applies lightweight dual-pooling spatial attention before each detection head to suppress background interference. Third, the Anisotropic Minimum Point Distance IoU (A-MPDIoU) loss normalizes horizontal and vertical coordinate deviations by the ground-truth width and height to improve localization of boxes with unequal scales. Experiments on 4325 images show that the proposed model achieves 94.25% mAP@50, 93.34% precision, 92.11% recall, and 92.72% F1-score, with 7.82 M parameters, 23.4 G FLOPs, and 95.21 FPS. Controlled ablation, sensitivity, and replacement experiments support the complementary contributions of the proposed components. Further validation on independently collected datasets is required.