LRFL-YOLO: A Large Receptive Field and Lightweight Model for Small Object Detection

Issues, such as a dilemma regarding the size of the receptive field, a substantial obstacle in background confusion, and a tradeoff between accuracy and computational efficiency, make tasks for small object detection in remote sensing challengeable. Thereby, an efficient and lightweight detector, called large receptive field and lightweight YOLO (LRFL-YOLO) based on YOLOv11 is proposed in this article, which includes four efficient modules to enlarge the receptive field: a dual-branch fusion attention (DBFA) module, a scale-aware extension of the unified IoU (SA-UIoU), enhanced C3k2 module incorporating wavelet transform convolution (WTC3k2), and an efficient upsampling convolution block (EUCB). These four modules enable our model to expand the receptive field while preserving local information and to deliver both high accuracy and computational efficiency. Experimental results on the DOTAv1.5 dataset, the RSOD dataset, and the VEDAI dataset verify that the LRFL-YOLO reaches the accuracy of 78.7%, 93.6%, and 74.5% in mAP50, outperforming several baseline models and other state-of-the-art (SOTA) methods. Moreover, LRFL-YOLO achieves improved computational efficiency with 2.4 M parameters (16% less than baseline model) and 6.3 GFLOPs (1.5% lower than baseline model), while maintaining real-time performance at 127 frames per second (FPS) on DOTA, 146 FPS on RSOD, and 147 FPS on VEDAI. These results represent a 9.48%, 16.8%, and 9.7% increase in FPS compared to the baseline, demonstrating a reduction in processing time relative to other benchmark models. Given the improvements above, this model is well suited for small object detection.