Pointer Defect Detection Based on Transfer Learning and Improved Cascade-RCNN.

Area of Science:

• Computer Vision
• Machine Learning
• Materials Science

Background:

• Detecting defects on pointer surfaces presents challenges due to internal variations and subtle feature characteristics.
• Existing methods struggle with accurate identification and classification of diverse pointer defects.

Purpose of the Study:

• To develop an advanced deep learning algorithm for precise detection of various pointer surface defects.
• To address limitations in current defect detection methods, particularly for subtle or internally varied defects.

Main Methods:

• Proposed a Transfer Learning and Improved Cascade-RCNN deep neural network (TICNET) algorithm.
• Enhanced ResNet-50 feature extraction using deformable convolutions to better learn pointer surface defect features.
• Integrated Online Hard Example Mining (OHEM) into Cascade-RCNN for improved defect classification.
• Utilized transfer learning by pre-training on a common defect dataset and fine-tuning on a specific pointer defect dataset.

Main Results:

• Achieved a 0.933 detection rate.
• Obtained a 0.873 mean average precision at an intersection over union threshold of 0.5.
• Demonstrated high-precision detection capabilities for pointer surface defects.

Conclusions:

• The proposed TICNET algorithm effectively detects pointer surface defects with high accuracy.
• The combination of deformable convolutions, OHEM, and transfer learning significantly enhances detection performance.
• TICNET offers a robust solution for practical defect detection needs in pointer manufacturing and inspection.