ESF-YOLO: an accurate and universal object detector based on neural networks

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• YOLOv5 is a proficient single-stage object detector widely used in industry.
• Existing YOLOv5 models have design limitations impacting performance, particularly with occluded objects and varying target scales.

Purpose of the Study:

• To enhance the performance of YOLOv5 for object detection tasks.
• To address limitations in feature learning, occlusion handling, and sample imbalance in existing models.
• To develop an improved object detection model, Efficient Scale Fusion YOLO (ESF-YOLO).

Main Methods:

• Introduced the Multi-Sampling Conv Module (MSCM) for enhanced low-level feature learning via multi-scale receptive fields and cross-scale fusion.
• Developed a Block-wise Channel Attention Module (BCAM) to improve the handling of occluded objects by weighting critical information channels.
• Designed a lightweight Decoupled Head (LD-Head) and redesigned the loss function to mitigate label-confidence asynchrony and sample imbalance.
• Proposed an adaptive scale factor for Intersection over Union (IoU) calculation to better accommodate diverse target sizes.

Main Results:

• ESF-YOLO demonstrated significant improvements on the SODA10M and CBIA8K datasets.
• Achieved increases of 3.93% and 2.24% in Average Precision at 0.50 IoU (AP50).
• Showcased gains of 4.77% and 4.85% in Average Precision at 0.75 IoU (AP75), and 4% and 5.39% in mean Average Precision (mAP).

Conclusions:

• ESF-YOLO effectively addresses key limitations of previous YOLOv5 models.
• The proposed enhancements lead to superior object detection performance, particularly in challenging scenarios like occlusion.
• ESF-YOLO exhibits broad applicability and improved accuracy for industrial object detection tasks.