A method for detecting metal surface defects using a dynamic fine-grained multi-branch encoder.

Area of Science:

• Computer Vision
• Materials Science
• Artificial Intelligence

Background:

• Small-scale metallic surface defect detection faces challenges in accuracy, feature distinction, and real-time processing.
• Existing methods struggle with precise localization and identifying subtle defects on metallic surfaces.

Purpose of the Study:

• To propose a novel detection framework, FF-MDE (fine-grained multi-branch diversified encoder), to overcome limitations in metallic surface defect detection.
• To enhance detection capabilities for improved accuracy, feature discrimination, and real-time performance.

Main Methods:

• Developed a direction-aware multi-branch convolutional block (BBDES) for orientation-sensitive feature extraction.
• Implemented a multi-scale fusion network (MFFN) with adaptive receptive field expansion and cross-resolution feature alignment.
• Utilized lightweight attention-guided up-sampling for improved fine-detail recovery and noise suppression.

Main Results:

• FF-MDE achieved high performance on NEU-DET and GC10-DET datasets, with mAP50 scores of 70.3% and 57.9% respectively.
• Demonstrated significant improvements over benchmark and existing methods, exceeding them by 4.1-7.6%.
• Achieved real-time inference speeds exceeding 60 FPS, proving its deployability.

Conclusions:

• FF-MDE offers a powerful and deployable solution for high-precision defect detection in industrial vision inspection.
• The proposed framework effectively addresses challenges in localization accuracy, feature discrimination, and real-time constraints.
• This advancement contributes to more robust and efficient quality control in metallic manufacturing.