Abstract:
This paper introduces a novel generative model based on an encoder-decoder architecture for defect detection within Industry 4.0 frameworks, focusing on the escalating need for automated quality control in manufacturing settings. Precision and efficiency, crucial in such environments, are significantly enhanced by our approach. At the core of our methodology is the strategic incorporation of random Gaussian noise early in the image processing sequence. This deliberate interference disrupts the model's ability to reconstruct images of defective parts, thereby enhancing both the accuracy and robustness of defect detection. The model further integrates skip connections during the decoding phase, with a special emphasis on the first two connections. These are augmented with multi-head attention mechanisms and spatial reduction techniques, followed by targeted convolutions. This intricate configuration helps preserve vital local features while filtering out superfluous data, facilitating precise image reconstruction
and effectively addressing the often problematic issue of locality loss during the upsampling process. Moreover, our model excels in maintaining contextual integrity and capturing multi-scale features, which is crucial for detailed defect detection. Each block of the architecture connects to a scaled version of the original image, allowing for nuanced feature analysis. Extensive testing and validation on real-world datasets have proven the model's high efficiency and accuracy in identifying defects, marking a significant advancement in automated quality control systems.