YOLO-SAD: Enhancing Small Arthropod Target Detection for Autonomous Port Inspection Robots

In port trade, the transport of foreign arthropods is often encountered, and manual inspection is usually needed. However, owing to the small size of arthropod targets and the interference of the background, detecting them is difficult. To address this challenge, this article introduces the YOLO-SAD network based on YOLOv11, which uses an inspection robot equipped with a camera and this network for target detection instead of manual inspection. In this study, a dataset containing 1640 images with 30,826 labeled instances of multiple common small-target arthropods and background was constructed. Afterward, a small arthropod detection network (YOLO-SAD) was designed. By constructing a spatial-to-channel dense convolutional fusion (SPDCF) module to replace part of the traditional convolutions, a CBAM-enhanced kernel convolution (CKC) module was used to replace the C3K2 module of the neck network, and a fully connected FPN (FcFPN) network structure was constructed to replace the original neck network structure. Experiments were conducted based on the constructed dataset, and the results revealed that YOLO-SAD, with 22.8% fewer parameters than YOLOv11 had, achieved a 4.2% higher precision, a 1.2% higher mAP@0.5, and a 2% higher mAP@0.5:0.95 than YOLOv11 did. Moreover, it performed well with the VisDrone dataset, demonstrating the strong performance and robustness of YOLO-SAD for small-target detection. This method has the potential for increased target detection accuracy and efficiency in port inspection robots for detecting foreign arthropods.