Boosting lightweight object detection with enhanced feature fusion and optimized receptive field

The average performance improvement using an input image resolution of 416 between the five most recent nano architectures of the You Only Look Once (YOLO) models is 0.3%, with the largest being 0.8% between YOLOv12-n and YOLOv11-n. Practical applications typically require real-time inference on edge devices, making it particularly challenging and important to improve the accuracy of lightweight object detectors. In this paper, we show how this could be achieved through efficient feature fusion and receptive field expansion. We propose two modules that improve the mean Average Precision (mAP) by up to 4%, while having a small computational impact. The first block is called the Hybrid Pooling Convolutional Squeeze-and-Excitation (HP-CSE), and it improves the inter-channel relationship between features. The second block is called the Depthwise Double-Stage Receptive Field Block (D2S-RFB) and it expands the receptive field. We introduce a methodology for integrating the two modules into the architectures of lightweight object detectors, and we prove the cross-architecture generalization by integrating them into the latest six models in the YOLO series. We train the models on the MS COCO dataset and evaluate their real-time performance on the NVIDIA Jetson Nano. For example, when adapting YOLOv7-tiny and YOLOv12-n with our HP-CSE and D2S-RFB, the mAP improves by 3.6%, 1.7% respectively, while achieving 8–9 frames per second on the NVIDIA Jetson Nano. The code and models are available at https://github.com/dbacea/yolo-tiny-hp-cse-d2s-rfb .