基于Swin Transformer轻量化的TFT-LCD面板缺陷分类算法

A lightweight deep learning model for TFT-LCD circuits defect classification based on swin transformer

在TFT-LCD面板缺陷检测中,检测对象背景复杂、缺陷细微且种类繁多,而工业生产实时性要求高,传统的缺陷分类算法往往难以兼顾精度和速度要求,无法适用于实际生产应用。为均衡TFT-LCD面板缺陷分类的准确率和速率,提出一种基于Swin Transformer的轻量化深度学习图像分类模型。首先对模型每层输入的特征图进行Token融合以减少模型计算量,从而提高模型的轻量化水平。其次引入深度可分离卷积模块以帮助模型增加卷积归纳偏置,从而缓解模型对海量数据的依赖问题。最后使用知识蒸馏方法来克服模型轻量化导致的检测精度下降问题。在自制TFT-LCD面板缺陷分类数据集上的实验表明,本文提出的改进模型相比基线模型,FLOPs计算量降低了2.6 G,速度指标提升了17%,而Top-1 Acc精度仅损失1.3%,且与其他图像分类主流模型相比,在自制数据集和公开数据集上都具有更均衡的精度和速度。

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.