The detection of the impurity rate in machine-picked seed cotton is crucial for precision agriculture. This study proposes a novel Cotton-YOLO-Seg cotton-impurity instance segmentation algorithm based on the you only look once version 8 small segmentation model (Yolov8s-Seg). The algorithm achieves precise pixel-level segmentation of cotton and impurities in seed cotton images and establishes a detection model for the impurity rate, enabling accurate detection of the impurity rate in machine-picked cotton. The proposed algorithm removes the Pyramid 4 (P4) feature layer and incorporates Multi-Scale Convolutional Block Attention (MSCBCA) that integrates the Convolutional Block Attention Module (CBAM) and Multi-Scale Convolutional Attention (MSCA) into the Faster Implementation of Cross Stage Partial Bottleneck with 2 Convolutions (C2f) module of the feature extraction network, forming a novel C2f_MSCBCA module. The SlimNeck structure is introduced in the feature fusion network by replacing the P4 feature layer with the small-target detection layer Pyramid 2 (P2). Additionally, transfer learning is employed using the Common Objects in Context (COCO) instance segmentation dataset. The analysis of 100 groups of cotton image samples shows that the Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) for impurity rate detection are 0.29%, 0.33%, and 3.70%, respectively, which are reduced by 52.46%, 48.44%, and 53.75% compared to the Yolov8s-seg model. The Precision (P), Recall (R), and mean Average Precision at an intersection over union of 0.5 (mAP@0.5) are 85.4%, 78.4%, and 80.8%, respectively, which are improved by 4.2%, 6.2%, and 6.4% compared to Yolov8s-seg model, significantly enhancing the segmentation performance of minor impurities. The Cotton-YOLO-Seg model demonstrates practical significance for precisely detecting the impurity rate in machine-picked seed cotton.

This study considers multi-class visual detection for processing tomato sorting under controlled laboratory conditions. In densely arranged images, occlusion and visual similarity can weaken the local boundary and texture cues needed to distinguish ripe tomatoes, unripe tomatoes, defective tomatoes, and soil clods. We propose SG-RTDETR, an RT-DETRv2 adaptation that combines detail-preserving downsampling, saliency-guided token encoding with residual spatial refill, context-aware feature organization, and adaptive cross-scale fusion. A four-class dataset was constructed from 1000 multi-object images and 400 single-object images used for supplementary representation learning. Across three independent training runs under controlled laboratory evaluation, SG-RTDETR achieved $87.8\pm 0.4$% ${\mathrm{mAP}}_{50:95}$, $92.4\pm 0.4$% ${\mathrm{mAP}}_{50}$, and $95.5\pm 0.3$% ${\mathrm{mAR}}_{50:95}$ (mean ± sample standard deviation). Relative to RT-DETRv2, the mean ${\mathrm{mAP}}_{50:95}$ increased by 3.4 percentage points, while total FLOPs remained at 61.17 G and forward-pass inference throughput decreased from 110.5 to 103.6 FPS. These results indicate an accuracy-oriented trade-off under the laboratory evaluation protocol; validation under realistic postharvest sorting conditions remains necessary.

Processing tomato harvesters require real-time sorting of ripe tomatoes, unripe tomatoes, defective tomatoes, and soil clods in dense material flows. The key challenge is that category evidence is subtle, local, and spatially uneven: defect textures, fruit boundaries, and soil-clod cues are easily diluted by adjacent objects and background clutter. To address this, we propose SG-RTDETR, which adapts RT-DETRv2 to selectively allocate representation capacity to informative target regions while retaining local evidence. The model preserves high-resolution details during downsampling, applies saliency-guided token selection to focus global encoding on target-related locations, and reintegrates encoded tokens through residual spatial refill to maintain feature continuity. Context-aware feature organization and spatially adaptive cross-scale fusion further reduce fragmented responses and background-dominated fusion. Together, these operations target two main error sources: lost local cues and cluttered cross-region interactions. We constructed a four-category processing tomato dataset with 1,000 multi-object images and 400 single-object images for supplementary representation learning. Under a unified evaluation protocol, SG-RTDETR achieved 87.9% mAP50:95, 92.5% mAP50, and 95.6% mAR50:95, outperforming RT-DETRv2 by 3.4, 3.6, and 0.6 percentage points, respectively, while maintaining 103.6 FPS. These results show that selective modeling of spatially uneven evidence improves real-time tomato sorting.