SAN-RNI: A Deep Learning Based Approach to Detect Anomalies on Reflecting and Non-Reflecting Surfaces

Detecting anomalies such as bumps and dents on reflective and non-reflective surfaces like vehicle surfaces is difficult due to specular noise, diffuse lighting, and unpredictable reflections. Traditionally, convolutional neural networks (CNNs) are used for this task; however, they suffer from limitations such as overfitting, high dependency on labelled data, and difficulty in detecting subtle or complex anomalies. CNNs often fail to extract and isolate high-level semantic features, leading to poor performance in critical industrial quality control applications. A potential substitute for CNNs is vision transformers (ViTs), offering better global context awareness and improved feature representation. Still, ViTs require large datasets for training and are computationally intensive, which limits their use in real-world industrial settings. To overcome these issues, we propose a hybrid algorithm combining MobileNet V3 and ResNet-34, enhanced with a sequential attention network (SAN-RNI). This model integrates deflectometry-based information with channel-wise attention mechanisms to better detect color and texture anomalies. By emphasising important areas and reducing extraneous background information, the attention layers increase accuracy and resilience. Our approach outperforms traditional CNN-based models in terms of accuracy and loss, as demonstrated by experimental findings on the MVTec AD and deflectometry datasets. This indicates the method's potential for dependable automated visual inspection in industrial settings.