A Fast Inference Vision Transformer for Automatic Pavement Image Classification and Its Visual Interpretation Method

Traditional automatic pavement distress detection methods using convolutional neural networks (CNNs) require a great deal of time and resources for computing are poor in terms interpretability. Therefore, inspired by the successful application Transformer architecture natural language processing (NLP) tasks, novel method called LeViT was introduced asphalt image classification. consists layers, transformer stages where Multi-layer Perception (MLP) multi-head self-attention blocks alternate residual connection, two classifier heads. To conduct proposed methods, three different sources datasets pre-trained weights based on ImageNet were attained. The performance model compared with six state-of-the-art (SOTA) deep learning models. All them trained transfer strategy. Compared to tested SOTA has less than 1/8 parameters original Vision (ViT) 1/2 ResNet InceptionNet. Experimental results show that after training 100 epochs 16 batch-size, acquired 91.56% accuracy, 91.72% precision, recall, 91.45% F1-score Chinese dataset 99.17% 99.19% German dataset, which is best among all Moreover, it shows superiority inference speed (86 ms/step), approximately 25% ViT 80% some prevailing CNN-based models, including DenseNet, VGG, ResNet. Overall, can achieve competitive fewer computation costs. In addition, visualization combining Grad-CAM Attention Rollout analyze classification explore what been learned every MLP attention block LeViT, improved interpretability model.