Thrips detection on sticky traps based on improved YOLOv11n

Thrips on sticky traps are extremely small and densely distributed, making automated detection challenging. This study proposes YOLOv11n-WNE, an improved YOLOv11n-based detector tailored for thrips detection in greenhouse environments. Three modifications are introduced: (1) a Wavelet Unpooling StarFusion (WSF) network in the neck, combining WaveletUnPool upsampling—which uses the discrete wavelet transform to preserve spatial details—with StarFusion feature enhancement that captures high-order multi-scale interactions through element-wise multiplication; (2) a Normalized Wasserstein Distance (NWD) loss that models bounding boxes as 2-D Gaussian distributions, providing smooth gradient responses for tiny-object localization; and (3) an EfficientHead detection head using grouped convolutions (G = 8) to compress parameters and computation. A dataset of 2, 400 images with 40, 480 annotated thrips instances was constructed for training (1, 680), validation (360), and testing (360). YOLOv11n-WNE achieves [email protected] of 91.5%, with precision of 88.2% and recall of 85.0%—improvements of 4.1, 4.5, and 5.1 percentage points over the baseline YOLOv11n, respectively—while reducing parameters to 2.11M (81.8% of baseline) and model size to 4.5 MB. On the GPU, the model runs at 123.8 FPS at an input size of 640×640, supporting its deployment for edge-based intelligent thrips monitoring in greenhouses. Beyond individual novelty, the contribution lies in problem-driven integration. WSF, NWD, and EfficientHead act on complementary stages of the detection pipeline—feature aggregation, localization supervision, and model compression, respectively. Together they address the failure modes of thrips-on-sticky-trap imagery: dense tiny objects, weak textures, and frequent overlap with non-target insects.