Vol. 54, Issue 4, pp. 567-579 (2024)

Abstract

To process the massive optical image data collected in machine vision systems and address the limitations of current learning detection models for real-time processing, this paper proposes a lightweight and real-time detection model based on YOLOX-Nano. While YOLOX-Nano is a lightweight object detection model, its detection accuracy is relatively low. Thus, this paper focuses on ensuring a lightweight model while maintaining high accuracy. The improved model incorporates an attention mechanism based on spatial and channel features to enhance the feature extraction capability of the YOLOX-Nano model. Additionally, a dual decoupled feature fusion approach is introduced to further improve the weighted fusion of feature maps extracted at different levels. This approach addresses the issue of smaller objects being overlooked in multi-object detection and enhances detection accuracy. Compared with the YOLOX-Nano baseline model, the proposed model achieves a detection speed of 59.52 FPS (frames per second) while increasing the AP50:95 metric. It meets the requirements for real-time detection, which is suitable for deployment on embedded systems, enabling the requirements of miniaturized optical processing tasks.