TLD-YOLO: A Lightweight Detector with Cross-Scale Adaptive Fusion and Dual Attention for Tiny Litchi Detection and Counting in Unstructured Orchards

Abstract

Accurate detection and counting of tiny litchi fruits in natural orchards are fundamental for approximate yield estimation and precision harvesting; however, small-target signals are easily lost during downsampling, while dense clusters, occlusion, and complex backgrounds further impair localization. Attention applied only after feature refinement cannot protect weak intermediate responses, while channel concatenation treats spatial details and semantics equally. This study proposes Tiny Litchi Detection YOLO (TLD-YOLO), a lightweight detector for tiny litchi detection and counting in unstructured orchards. First, customized Embedded Cross-Scale Attention (ECS Attention) is embedded mid-stream within the backbone refinement process and incorporates support-constrained local-contrast enhancement to protect weak fruit responses. Then, the Cross-Scale Adaptive Fusion Module (CAF Module) replaces five concatenation nodes in the main Neck pathway with scalar-weighted fusion and adds local-detail compensation at the high-resolution P4–P3 stage, reducing parameters while preserving detail. Finally, optimized Directional Spatial Attention (DS Attention) uses gated average–maximum directional descriptors to recalibrate refined features, enhancing fruit regions and suppressing backgrounds. Inner Minimum Point Distance Intersection over Union (Inner-MDPIoU) with auxiliary box scaling further improves fine localization. Together, these components form a unified lightweight pipeline tailored to dense tiny-fruit detection. TLD-YOLO achieves 88.0% mean average precision at an intersection-over-union threshold of 0.5 (mAP@0.5) and 60.2% mAP@0.5:0.95 with 7.84 M parameters. Compared with the baseline, the two mAP metrics improve by 2.0 and 2.6 percentage points with 18.3% fewer parameters. In video-based visible-fruit counting, missed detections and false positives in the respective maximum-count frames are 36.7% and 66.7% lower than for the baseline. The view-dependent maximum count does not measure the complete three-dimensional canopy fruit load but can provide a simple approximate indicator for preliminary yield estimation. TLD-YOLO therefore provides a practical reference for tiny litchi detection and counting with potential application to approximate orchard yield assessment.