The problem is what the researchers call adhered leaf clusters, or ALCs. During handling, fresh tea leaves become structurally entangled, forming clumps such as two one-bud-one-leaf units stuck together, or a one-bud-two-leaf unit adhered to a multi-leaf unit. When an RGB camera on a conveyor belt images such a clump, the algorithm sees what looks like a single leaf unit, and the misjudgment rate climbs with the number of clusters present. Mechanical countermeasures such as S-shaped soft rakes and vibration transmission have been tried, but their effectiveness falls short of industrial requirements. Writing in Smart Agricultural Technology, Jian Wang of the tea processing research group and colleagues describe a reworked neural network, dubbed SDSI-YOLO, that learns to see through the tangle.

The baseline for their work is YOLOv8n, the nano variant of the widely used You Only Look Once object detection family. YOLO detectors are prized in agriculture because they run fast enough for real-time sorting on affordable hardware. But the standard architecture has structural weaknesses for this task. Its CSPDarknet backbone progressively expands the receptive field while shrinking feature map dimensions through repeated convolution and pooling operations. Small targets — and adhered leaf clusters generate intricate small-scale details such as occluded bud tips, overlapping edges, and shadowed regions — lose their edges and textures along the way, leaving high-level feature maps with insufficient discriminative information. Meanwhile, in the PAN-FPN neck that fuses multi-scale features, small target signals tend to be diluted by dominant large target features, impairing accuracy.

The team’s answer is a triple-driven redesign targeting exactly these weaknesses. First, they replaced the CSPDarknet backbone with StarNet, a lightweight network built around the so-called Star Operation. In a StarBlock module, element-wise multiplication across channels establishes second-order feature interactions, which the researchers say amplifies subtle texture differences between adjacent adhered leaves — a capability absent from conventional first-order convolution. For clusters with blurred adhesive boundaries, this mechanism extracts more discriminative boundary features rather than merging overlapping regions into one unified representation, reducing the chance that multiple stuck-together leaves are misread as a single target.

Second, the fixed upsampling layers in the neck were swapped for DySample, which reformulates upsampling as point-wise adaptive sampling. Instead of resampling features at fixed intervals, DySample predicts learnable spatial offsets for each target location, directing sampling points toward regions with clear texture and strong semantics. In the original YOLOv8n, fixed-interval nearest-neighbor upsampling tended to concentrate excessive sampling points on ambiguous adhesion boundaries, distorting feature reconstruction and losing edge detail. DySample’s dynamic resampling reduced feature loss at adhesion junctions during multi-scale fusion and improved boundary recall — all while adding essentially no computational cost.

Third, the bounding box regression loss was upgraded to a composite Shape-Inner-IoU function. Most adhered leaf clusters are long-narrow targets with large aspect ratios, yet traditional IoU losses apply equal penalties to horizontal and vertical deviations, ignoring target morphology. Shape-IoU incorporates direction-aware penalty terms that dynamically adjust the weights of positional and dimensional deviations based on each object’s aspect ratio, focusing the model on length-direction deviations of narrow leaves. Inner-IoU applies scaled auxiliary bounding boxes to balance gradient strength between high- and low-IoU samples, preventing easy samples from dominating training while hard adhesion cases are neglected. Together, these mechanisms markedly refined localization of the irregular clusters.

The evidence base is substantial. The team built a custom acquisition system with a 22-megapixel Hikvision industrial camera, dual LED strip light sources, and a conveyor belt running at 6 centimeters per second, collecting 550 sample images with 7,520 manually annotated bounding boxes. After augmentation with random rotation, flipping, and saturation adjustment, the dataset grew to 10,710 images, partitioned 7:2:1 into training, validation, and test sets. Six batches of samples were gathered across the April-to-November picking period, with images acquired within 36 hours of harvesting to guarantee freshness. All experiments were independently repeated five times with different random seeds.

The module-by-module comparisons were revealing. Among five lightweight backbones tested — EfficientViT, MobileNetv4, FasterNet, ConvNeXtv2, and StarNet — StarNet achieved the highest F1-score of 88.9 along with the smallest model size, 3.86 MB, and the lowest computational load, 5.2 gigaflops. Among four upsampling methods, DySample delivered results comparable to the best competitor at substantially lower cost, matching the baseline’s 2.1-millisecond inference time. Among five IoU loss functions, the composite Shape-Inner-IoU achieved the highest mean average precision of 91.8 percent for cluster detection. Ablation experiments then confirmed the payoff of stacking all three: the full SDSI-YOLO reached a mean average precision of 92.6 percent and an F1-score of 89.2, outperforming RT-DETR and ten classic YOLO models, while shrinking to just 3.89 MB — roughly a third smaller than the baseline — and inferring at 3.1 milliseconds per frame.

Perhaps most convincing is the visualization evidence. Using Gradient-weighted Class Activation Mapping, the team compared where the original and improved models direct their attention. On adhered cluster samples, the original YOLOv8n formed wide rectangular activation areas covering entire clumps, failing to distinguish adhesion junctions, and its attention became scattered and blurred on the most complex multi-leaf composites. SDSI-YOLO, by contrast, directed attention to the overlapping boundaries between tea buds and leaves and clearly resolved the outlines of individual leaf units even in tangled composites — confirming that StarNet genuinely enhanced perception of subtle texture differences at adhesion positions rather than merely shifting statistics.

The implications extend past tea. The researchers note that no prior study has combined RGB imaging with an improved YOLO approach for adhered leaf cluster detection, and their framework — a detail-preserving backbone, adaptive upsampling, and shape-aware loss — addresses the general challenge of detecting severely occluded, small, irregular objects on fast-moving belts. For the tea industry, the payoff is concrete: a feasible, lightweight solution for industrial sorting of fresh green tea leaves that keeps high-grade one-bud and two-bud products out of low-grade streams, with composite clusters routed to secondary screening to minimize losses. For anyone watching the intersection of computer vision and food processing, SDSI-YOLO is a compelling demonstration that sometimes the smartest fix is not a bigger model, but a better-targeted one.

Subject of Research: Enhanced YOLOv8n for adhered leaf cluster detection in RGB-based fresh green tea leaf sorting

Article Title: Enhanced YOLOv8n for adhered leaf cluster detection in RGB-based fresh green tea leaf sorting

Article References: Wang, J., Liu, Y.-J., Lin, Y.-Y., Zhang, C.-C., Liu, S.-N., Ji, B., Wang, Q., Du, L.-H., Zheng, L., & Ma, F. (2026). Enhanced YOLOv8n for adhered leaf cluster detection in RGB-based fresh green tea leaf sorting. Smart Agricultural Technology, 15, Article 102640. https://doi.org/10.1016/j.atech.2026.102640

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102640

Keywords: Enhanced, YOLOv8n, adhered, leaf, cluster, detection, RGB-based, fresh, green, sorting, scientific research