ADM-Net: attentional-deconvolution module-based net for noise-coupled traffic sign recognition

Convolutional Neural Networks (CNNs) have become primary technologies in computer vision systems across multiple fields. Its central characteristic is to slide filters on input images and repeats the same procedures to obtain the image’s robust features. However, conventional CNNs struggle to classify objects when the input images are contaminated by unavoidable external noises such as missing information, blur, or illumination. This paper proposes an attentional-deconvolution module (ADM)-based net(ADM-Net) in which ADMs, convolutional-pooling, and a fully convolutional network (FCN) are applied to improve classification under such harsh conditions. The structure of ADM includes an attention layer, deconvolution layer and max-pooling. The attention layer and convolutional pooling help the proposed network maintain key features through convolution procedures under noise-coupled environments. The deconvolution layers and fully convolutional structure have advantages in providing additional information from upscale feature maps and enabling the network to store local pixel information. The ADM-Net was demonstrated on the German traffic sign recognition benchmark with different noise cases comparing densenet, multi-scale CNN, a committee of CNN, hierarchical CNN, and a multi-column deep neural network. Demonstrations of ADM-Net achieve the highest records in different cases such as 1) blur and missing information case: 86.637%, 2) missing information and illumination case: 92.329%, and 3) blur, missing information, and illumination case: 80.221%. Training datasets for ADM-Net have limited conditions, the proposed network demonstrates its robustness effectively under noise-coupled environments.