This thesis introduces novel automated visual defect detection approaches that effectively address the challenges of data scarcity and imbalance. In the manufacturing industry, conventional defect detection systems rely on end-to-end supervised learning methods that necessitate abundant labeled data, including defective samples. However, such data is often insufficiently available. In light of this, we propose two alternative approaches. The first approach combines unsupervised learning anomaly detection with rule-based computer vision, enabling effective defect detection with a smaller dataset consisting exclusively of non-defective samples. The second approach leverages rule-based computer vision exclusively, eliminating the need for any training data.

To demonstrate the practicality and efficacy of the proposed approaches, this study uses the case study of Automated Fiber Placement (AFP) and design, implement, and evaluate both methods for defect detection in this industry. Specifically, these methods are tested on depth map images of the composite surface obtained using Optical Coherence Tomography (OCT) technology. Before utilizing these images for defect detection, certain preprocessing steps, such as noise filtering, are applied to enhance their quality.

In the anomaly detection approach, the process begins with utilizing Hough Transform to estimate the boundaries of each composite strip (tow). Subsequently, a sliding window traverses along each tow, extracting small patches. A subset of these patches that are free from anomalies is used to train the autoencoder. Since the autoencoder is trained using normal samples, it can generate more precise reconstructions for these patches compared to abnormal ones. Consequently, the reconstruction error value serves as a quantitative metric to determine the presence of potential anomalies within each patch. By aggregating these values, an anomaly map is generated, enabling the identification of manufacturing defects within the depth map. The results demonstrate that despite the autoencoder being trained with a limited number of images, the proposed approach achieves satisfactory accuracy in binary classification and effectively localizes the defects.

The rule-based method proposed in this study effectively identifies gaps and overlaps, which are the most common manufacturing defects in AFP. This approach combines classical computer vision techniques to identify the outlines of individual tows, enabling a comparison between consecutive tows to identify any potential gaps or overlaps. To assess the effectiveness of the proposed approach, this study compares the detected defects with ground truth annotations provided by human experts. The results affirm a high accuracy in segmenting gaps and overlaps.