Real-time object detection via enhanced YOLOv5s with attention mechanisms and multi-scale feature fusion

This study aims to enhance the efficiency of automation and robotic technologies by developing an advanced computer vision model for real-time object detection and recognition. The model is designed to address challenges such as complex backgrounds and small object detection in various environments. We propose an improved YOLOv5s-based model incorporating a spatial and channel convolutional attention mechanism within the BottleneckCSP structure to emphasize essential features and mitigate background noise. An enhanced Bi-directional Feature Pyramid Network (BiFPN) is integrated to improve multi-scale feature fusion, while the Kmeans++ algorithm is employed for re-clustering prior boxes to enhance object localization. Additionally, lightweight GhostC3 and GhostConv modules are developed to reduce model size without compromising accuracy. Empirical evaluations demonstrate that our model achieves an average detection accuracy of 98.5% and a recall rate of 90.6%, surpassing existing state-of-the-art methods. The model processes frames at a rate of 41 frames per second, meeting the stringent real-time detection requirements. Compared to baseline YOLOv5s, our model achieves a 2.9% improvement in accuracy and a 6.1% improvement in mAP while reducing model parameters by 8.0% (from 7.5 to 6.9 M), demonstrating superior balance between accuracy, speed, and model size for resource-constrained environments. Ablation studies confirm the effectiveness of each proposed improvement, with the full model outperforming the baseline YOLOv5s in terms of accuracy, mAP, and recall while maintaining computational efficiency. The proposed model significantly enhances real-time object detection capabilities through optimized feature extraction and lightweight design.