Based on Improved YOLOv7 Small Target Detection

Abstract

In recent years, deep learning-based object detection has developed rapidly, but its performance in the small object detection field is not ideal compared to the natural scene image domain. Hence, this paper proposes an improved small object detection algorithm based on YOLOV7 algorithm. Firstly, based on Ghost model, the ELAN module in backbone structure is improved to Gmodel which effectively reduces computation and improves accuracy. Secondly, this paper introduces a Triplet Attention-improved small object attention module Amodel in YOLOV7's head structure; through Amodel's cross-latitude interaction function, it enhances the feature detection performance for small objects. Experiments were conducted on RSOD dataset and our method increased yolov7's AP50 by 1.65mAP and AP50-95 by 1.88mAP while also reducing FLOPs by 0.2G, making it more suitable for dense small target scenes for object detection.