Deep Learning for Automated Quality Inspection of Mechanical Parts : An Analysis of Image Representations for Lathed Hole Defect Detection

Sammanfattning

Quality control is an essential part of modern manufacturing, where reliable defect detection is necessary to ensure that produced components satisfy required standards.Traditional visual inspection is often manual, time-consuming, and sensitive to human error, which has increased interest in automated computer vision-based inspection systems. While deep learning methods have shown strong performance for surface defect detection, less attention has been given to defect detection in components with complex internal geometries. This thesis investigates automated defect detection in small lathed holes using deep learning-based image classification. Detecting defects inside holes introduces several domain-specific challenges, including shadows, occlusion, and varying defect depth, which can obscure important visual information. To address depth-related visibility variations, images captured at two focus depths were used. In addition, three region-of-interest preprocessing strategies were evaluated: whole-image, inside-region, and contour-only representations. Multiple convolutional neural network architectures, including VGG, ResNet, and EfficientNet, were trained using transfer learning in order to evaluate the effect of preprocessing and model architecture on defect detection performance. Both rule-based multifocus pipelines and multi-input deep learning configurations were investigated and compared. The results showed that the multi-input VGG19 model using the whole image representation achieved the strongest overall performance (0.85 F1-Score inindustrial testing). The study demonstrates the importance of preprocessing, validation and multi-focus information for defect detection in complex internal geometriesand contributes knowledge about deep learning-based inspection of lathed industrial components.