Feature-Based Transfer Learning for IoT-Enabled Defect Detection for Quality Control in Industrial Manufacturing Processes: A DenseNet Evaluation

Surface defects in industrial products can lead to performance issues or early failure. Automated optical inspection using computer vision and deep learning provides an efficient approach for quality control. This study investigated a feature-based transfer learning technique for classifying surface defects in images. The resampled KolektorSDD dataset containing 50 defect and 50 non-defect images of electrical commutator surfaces was utilised. Pre-trained DenseNet convolutional neural network (CNN) models were used to extract discriminative features from the images. A support vector machine (SVM) classifier was then trained on these features to classify images as either defective or non-defective. Three DenseNet architectures were tested, namely DenseNet121, DenseNet169, and DenseNet201. All three pipelines achieved 100% accuracy on the test and validation sets when used with the SVM classifier. Owing to the size and the overall performance of the pipeline, it was established that the DenseNet169-SVM was suitable for the given dataset. The high accuracy demonstrates that transfer learning can provide an effective approach for classifying surface defects compared to training a deep neural network from scratch. This allows for the development of automated quality control and surface inspection systems powered by computer vision and deep learning.