CSTAR-Det: A lightweight vision-based framework for conveyor-based cashew kernel quality assessment

Automated conveyor inspection of agricultural products requires fine-grained visual discrimination, stable temporal counting, and computational efficiency under variable acquisition conditions. This study presents CSTAR-Det, a lightweight framework for detecting, tracking, and counting cashew kernels by commercial quality category. The principal methodological contribution is the Cross-Stage Partial Triplet Adaptive Residual (CSTAR) block, which embeds cross-dimensional triplet attention in a partial dual-branch structure and couples multiplicative feature interaction, depthwise-separable refinement, and input-dependent residual modulation. The detector combines cross-scale feature aggregation, content-aware upsampling, and a two-scale end-to-end non-maximum suppression (NMS)-free prediction head. Bag of Tricks for SORT (BoT-SORT) association and one-time line-crossing registration extend frame-level detections to continuous category-wise counts. Evaluation uses 549 conveyor images containing 8605 annotated kernels in three classes. CSTAR-Det attains 96.4% mean average precision at an intersection-over-union threshold of 0.5 ([email protected]) and 81.0% averaged over thresholds from 0.5 to 0.95 ([email protected]:0.95), with 3.0M parameters, 17.4 GFLOPs, and 62.34 frames/s under the workstation detector-throughput protocol. Component-level ablations isolate individual CSTAR operations, with adaptive residual modulation producing the largest decrease in strict-IoU accuracy when removed. Across three 500-object conveyor sequences, the integrated BoT-SORT pipeline yields a mean absolute counting error of 2.33 objects. Controlled perturbation tests identify strong blur and viewpoint variation as the most influential degradations. The results support CSTAR-Det as a compact accuracy–complexity operating point for continuous conveyor inspection. The current evaluation is limited to a single processing site and imaging setup; independent cross-site validation is needed before deployment across different processing sites and imaging conditions.