Local Mixer with Prior Position for Cars’ Type Recognition

Abstract

Deep learning has attracted attention widely as the successful application of deep learning for vision tasks, such as image classification, object detection and so on. Due to the robustness and universality of deep learning, automotive manufacturing, a crucial part of national economy, needs deep learning to make production lines more intelligent and improve efficiency. However, some superior generally deep learning models, such as ViT, TNT, and Swin transformer, cannot meet automotive manufacturing requirements with high accuracy on a specific scene. As for automotive production lines, engineers usually adopt some smart designs, which can provide prior knowledge for designing deep learning models. Specifically, in an image, the position of target is usually fixed. Therefore, in order to take advantage of prior position, this paper designs a local mixer with prior position to capture local feature. Its main idea is that dividing the whole feature map into window feature maps and connecting window feature maps along channel dimension in order to make convolution kernel parameters for each window feature map are independent from others. Besides, MLP is adopted as global mixer to capture global feature and the pyramidal architecture with CNN is adopted. Comprehensive results demonstrate the effectiveness of proposed model on cars’ type recognition. In particular, the proposed model achieves 97.938% accuracy on our data set, surpassing some transformer-like models.