Research on Lightweight Small Object Detection Algorithm Based on Context Representation

Compared with general object detection, small object detection is more challenging due to its low resolution and limited anti-interference noise ability. Fully utilizing contextual semantic information is of great significance for solving the problems of small object detection. This paper proposes a lightweight framework algorithm based on a contextual semantic fusion model, which is built upon the YOLOv7 model. This framework model consists of three parts: a backbone network, a multi-scale feature representation network, and a detection head. Among them, partial convolutional (PConv) is utilized to construct a backbone network (P-Net), which ensures detection performance while further reducing computational complexity. The convolutional self-attention model is employed into traditional feature pyramid network (FPN) structures to reduce information loss during up-sampling and down-sampling processes, and a differential detection head is used to detect targets of different sizes. Comparative experiments on the aerial image tiny object detection (AI-TOD) dataset show that the proposed model achieves average precision of 21.6 and 161 frame/s on the AI-TOD benchmark respectively, surpassing other small detection models. In addition, the detection performance is superior to other traditional detection models under the condition of low parameter quantities and computational complexity. The results of ablation experiments indicate that the improved models proposed in this paper are effective.