A multi-scale small object detection algorithm SMA-YOLO for UAV remote sensing images

Abstract

Detecting small objects in complex remote sensing environments presents significant challenges, including insufficient extraction of local spatial information, rigid feature fusion, and limited global feature representation. In addition, improving model performance requires a delicate balance between improving accuracy and managing computational complexity. To address these challenges, we propose the SMA-YOLO algorithm. First, we introduce the Non-Semantic Sparse Attention (NSSA) mechanism in the backbone network, which efficiently extracts non-semantic features related to the task, thus improving the model's sensitivity to small objects. In the model's throat, we design a Bidirectional Multi-Branch Auxiliary Feature Pyramid Network (BIMA-FPN), which integrates high-level semantic information with low-level spatial details, improving small object detection while expanding multi-scale receptive fields. Finally, we incorporate a Channel-Space Feature Fusion Adaptive Head (CSFA-Head), which fully handles multi-scale features and adaptively handles consistency problems of different scales, further improving the robustness of the model in complex scenarios. Experimental results on the VisDrone2019 dataset show that SMA-YOLO achieves a 13% improvement in mAP compared to the baseline model, demonstrating exceptional adaptability in small object detection tasks for remote sensing imagery. These results provide valuable insights and new approaches to further advance research in this area.

Keywords: Feature fusion; Multi-branch auxiliary; Object detection; Remote sensing images.

Fig. 1

The overall structure of SMA-YOLO.

Fig. 2

Structure of NSSA, where represents , and and represent H/S and W/S respectively. The figure illustrates the case when the sparse coefficient S is 2, dividing the input features into 4 non-overlapping tensor blocks of different colors. When the sparse coefficient S is 1, it is equivalent to a single block, performing global attention.

Fig. 3

Structure of semantic and detail fusion modules.

Fig. 4

The Dynamic Fusion Weights Mechanism of CSFA-Head.

Fig. 5

The distribution of the objects in the dataset: (a) distribution of the number of classes; (b) the width and height distribution of the target; the concentration of the distribution is indicated by the color gradient from light white to dark blue, indicating that the distribution becomes more and more concentrated.

Fig. 6

Compare thermal maps generated for different model detection results: (a) Original images. (b) Heat maps of YOLOv8n. (c) Heat maps of SMA-YOLO.

Fig. 7

The visualization results between SMA-YOLO and YOLOv8n are compared on VisDrone2019, SSDD, and RSOD datasets.