Year: 2025
Type: article
Abstract: Object detection, as one of the core tasks of computer vision, possesses extensive implications in domains such as autonomous driving and video surveillance. Single stage detectors represented by the YOLO series have attracted attention for their efficient inference speed, but their detection accuracy still faces challenges in complex scenarios. The traditional YOLO model extracts feature through stacked convolutional layers, but lacks explicit modeling of inter channel dependencies, resulting in insufficient ability to distinguish key features. Attention mechanisms have significantly improved model performance by dynamically adjusting feature weights. The SE module proposed by SENet enhances the response of important channels through channel recalibration, but existing works often embed it directly into the Backbone network, which may cause redundant calculations due to module stacking and affect real-time performance. In addition, YOLO's Neck part is responsible for multi-scale feature fusion, and its optimization is crucial for detection performance, but there is little research on attention design for this region. This article proposes a lightweight SE-YOLO model that integrates the SE module into the Neck section of YOLO. Specifically, a channel attention mechanism is introduced after the SPP layer to capture channel level statistical information through global average pooling, and adaptive weights are generated using fully connected layers to enhance target related features while suppressing background noise. Experiments on the PASCAL VOC dataset show that the improved model performs well [email protected] By improving the indicators by 1%, This research offers fresh concepts for the lightweight improvement of object detection models.
Author: Yizhe Zheng
Institution: City University of Hong Kong
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4410899907