Few-Shot Defects Classification of Vehicular Display Screen Based on Improved DCGAN and Swin Transformer

Conventional defect classification approaches encounter significant challenges due to the coexistence of multi-type defects, complex backgrounds, and small-scale defect sample datasets, which hinder their ability to satisfy the stringent demands of modern manufacturing processes. To address the challenge of accurate multi-class defect classification for vehicular display screens under few-shot scenarios, this paper constructs a vehicular display screen defect datasets, and proposes an improved defect image generation model based on Deep Convolutional Generative Adversarial Network (DCGAN) and a Swin Transformer-based defect classification model for vehicular display screens. These two models are, respectively, used to achieve the generation of defect images of the vehicular display screen and the classification of defect images. First, the proposed method enhances the DCGAN architecture by incorporating a multi-scale self-calibration convolution block. This innovation strengthens the network capacity for hierarchical feature representation by fusing global contextual information with local spatial details, thereby improving the generation fidelity of defect samples with few data samples. Second, the improved Swin Transformer is augmented with a cross-attention block and a multi-pooling block. This modification enables adaptive cross-dimensional feature interaction and discriminative background feature extraction. Thus it significantly improves defect discriminability of the model in complex visual scenes. Extensive experiments demonstrate that the improved DCGAN model effectively alleviates the few samples limitation by generating high-fidelity synthetic defect images, and the improved Swin Transformer classification model achieves the state-of-the-art performance with a 99.8% classification accuracy on a multi-class defect dataset, fully satisfying the requirements of actual industrial production.