SSK-Yolo: Global Feature-Driven Small Object Detection Network for Images

Abstract

Timely and effective locust detection to prevent locust plagues is crucial for safeguarding agricultural production and ecological balance. However, under natural conditions, the “colour mixing mechanism” of locusts and the small scale of locusts in high-resolution images make it difficult to detect locusts. In this study, we propose a multi-scale prediction network SSK-Yolo based on YoloV5 to effectively solve the above two problems. Firstly, in the data preprocessing stage, in order to better adapt to the relatively small-scale targets, we use the k-means algorithm to cluster the a priori frames to obtain anchor frames with appropriate scale sizes. Secondly, in the backbone, we still use the traditional convolution to extract the shallow graphical features, and we use swin-transformer to extract the deep semantic features, so as to improve the accuracy of feature extraction and fusion for small targets in high-resolution images. In addition, in the data post-processing stage, we replace the NMS algorithm with the soft-nms algorithm by setting a Gaussian function for the neighbouring detection frames based on the overlapping part instead of suppressing all of them. A series of experimental results on the publicly available East Asian locust dataset demonstrate that SSK-Yolo outperforms YoloV5 with a 5% improvement in precision, 1.64% in recall, 12% in mAP, and 2.66% in F1-score. SSK-Yolo provides an efficient and viable solution for locust detection in the field of pest and disease control.