Image Classification of Printed Circuit Boards using Deep Learning

Sammanfattning

The increasing volume of electronic waste presents significant environmental challenges, with large amounts of unaccounted valuable materials and hazardous substances. To recover these materials and mitigate environmental harm it is essential to recycle and sort the integral components of electronic devices, namely Printed Circuit Boards. This thesis explores the potential of Deep Learning to identify Printed Circuit Boards using image classification. Traditionally, Convolutional Neural Networks have been the dominant choice for computer vision tasks due to their image-specific biases. However recently, Transformers have emerged demonstrating competitive performance. Based on self-attention mechanisms, Transformers can capture global patterns and long-range dependencies, outperforming Convolutional Neural Networks with sufficient data. Advances in hybrid architectures that blend the translational invariance and locality bias of Convolutional Neural Networks with the scalability and long-range dependency modeling of Transformers, have led to state-of-the-art performance in image classification. This thesis explores several Deep Learning models and demonstrates that ConvNeXt, a hybrid architecture, performs well in classifying Printed Circuit Boards. Utilizing Deep Learning models necessitates access to sufficient amounts of data. Due to the limited availability of annotated data, this thesis tackles the issue by leveraging a combination of transfer learning and data augmentation. These model enhancement methods can boost the predictive performance of ConvNeXt while maintaining computational efficiency. The findings reveal that CutMix outperforms the other data augmentation methods, achieving an accuracy of 93.5% with transfer learning. Additionally, training only the dense layer presents the optimal results as fine-tuning more layers did not result in noticeable improvements.