Defect detection in thin film transistor-liquid crystal display （TFT-LCD） circuits is a challenging task because of the complex background setting， different types of defects involved， and real-time detection requirements from industry. Traditional methods have difficulties in satisfying the dual requirements of detection speed and accuracy. To address this challenge， in this study， a deep learning method is developed for image classification based on the Swin Transformer technique. First， token merging is used to reduce the computational complexity of each layer of the model， thus improving computation efficiency. Then， a depthwise separable convolution module is introduced to add convolutional bias to reduce the reliance on massive data. Finally， a knowledge distillation method is applied to overcome the problem of reduced detection accuracy caused by the less-intensive computation design. Experimental results on the self-made dataset demonstrate that the proposed method achieves a 2.6 G FLOPs reduction and a 17% speed improvement compared to baseline models， with only a 1.3% Top-1 accuracy precision reduction. More importantly， the proposed model achieves better balance on accuracy and detection speed on both self-made and public datasets than existing mainstream models on image classification in the TFT-LCD manufacturing industry.