Lightweight precision model for watermelon seed group density estimation and counting

• This paper innovatively proposed a lightweight precision watermelon seed group density estimation and counting model-LOYOLO. • HGNetV2 is used as the backbone. GhostConv is applied in LightHGBlock to reduce parameters. • LSKA enhances focus by splitting large kernels, improving flexibility, and reducing parameters. • A group detection and counting method based on the LOYOLO model is originally proposed. • One dataset of watermelon seeds with different degrees of occlusion was first time generated. The accurate counting of overlapping watermelon seeds is a key foundation for seed quality testing, breeding selection, resource allocation, and other processes. To improve the counting accuracy for flat and slightly overlapping seeds, we introduce LOYOLO-GC, a Lightweight Occlusion YOLO8n-based group counting model. It adopts HGNetV2 as its backbone, where HGBlocks extract multi-level features for improved learning. GhostConv replaces the standard convolution in HGBlocks, forming LightHGBlock to reduce the number of parameters by generating intrinsic and ghost feature maps with fewer kernels. In addition, a Large Separable Kernel Attention mechanism (LSKA) is used to decompose deep convolution kernels into horizontal and vertical 1D kernels, enabling efficient large kernel attention with lower computational and memory cost. After optimizing the model, we build a multi-occlusion watermelon seed dataset and employ it to develop a LOYOLO-based group counting method. The experimental results show that LOYOLO-GC outperforms SOTA models, achieving 96.08 % accuracy and 86.66 % mAP, an improvement of 0.48 % and 1.67 %, respectively. The model parameters decrease by 63.8 % and GMACs decrease by 38.9 %. Counting accuracy is also improved, with ACC increasing by 5.32 % and L-ACC increasing by 5.04 %, while MAE and RMSE are decreased by 3.68 and 3.28, respectively.