A YOLO-ESAM Algorithm with Occlusion Immunity for Small Object Detection

Rice yield estimation is a crucial aspect of agricultural production, where the number of grains per panicle directly correlates with yield. Traditional manual counting methods are labor-intensive, inefficient, and costly. To address these challenges, this study proposes a YOLO-ESAM. rice grain counting and detection algorithm. The algorithm specifically addresses the challenges of small grain size, occlusion, and adhesion under natural growth conditions. Key improvements include: Replacing backbone modules with RepNCSPELAN4 integrated with the effectiveSE attention mechanism to enhance critical feature extraction while reducing computational redundancy and preserving valid features. Introducing the SEAM-Detect module to improve detection accuracy in occluded grain scenarios. Experimental results demonstrate performance improvements across YOLOv8 to YOLOv11 architectures. Notably, the enhanced YOLOv9 model achieves 96.7% precision, 90.1% recall, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.957 \% \text{mAP}50$</tex>, representing improvements of 2.0 percentage points, 1.92 percentage points, and 1.38 percentage points over baseline YOLOv9, respectively.