Detection of small objects in complex long-distance scenes based on Yolov3

Abstract

In view of the difficulty in detecting and managing illegal buildings, this article uses high-definition cameras mounted on a tower to regularly capture images to construct the “Suspected Illegal Building Information” dataset. Due to the high distance of the tower from the ground, most of the collected data images contian many small targets, and there are many types of target objects to be detected and target size is quite differences, resulting in low detection accuracy. Based on the above problems, this paper makes the following improvements based on the Yolov3 algorithm: (1) using K-menas re-clustering anchor boxes, (2) introducing CIoU optimization loss function, (3) adding 104×104 scale for feature extraction, (4) using Soft-NMS instead traditional NMS algorithm, in addition, the data set is enhanced and optimized. The experimental results show that, compared with the original Yolov3, the improved algorithm has improved the detection accuracy of this experimental data set by 19.1%, which can well meet the needs of the project.