Research on Classification Method of Film Damage Image Based on Improved ResNet50

Abstract In view of the challenges posed by traditional methods in accurately identifying the damage types of film images, this paper proposes an improved approach by leveraging transfer learning based on ResNet50, incorporating the channel attention mechanism from the Convolutional Block Attention Module (CBAM). Additionally, the AlphaDropout model is integrated with the SeLU activation function to enhance stability and performance, and the resultant model is named CBAM-ResNet50. The film dataset employed in this study comprises four types of damage: cracks, dewetting, particles, and scratches. Extensive training and testing of the CBAM-ResNet50 model on this dataset demonstrate significant improvements in classification accuracy. Compared to the original ResNet50 model, the proposed method achieves a remarkable 25.57% improvement in accuracy, reaching 90.58%. This work highlights the potential of combining attention mechanisms and advanced activation strategies to address complex image classification tasks. Furthermore, the approach paves the way for practical applications in quality inspection and automated defect detection in industrial processes