[C16] A Unified Embedding for Image Classification and Pixel-Level Regularization


Noise robust image classification is essential for successful deep learning deployment for the internet of things (IoTs) especially for the edge devices under stingy energy budget. Inherent image sensor noise and intentional image quality degradation for region of interest (RoI) coding reduce accuracy of image classification. In this paper, we enhance the accuracy of image classification on the perturbed images by utilizing embedding space for both image classification and additional pixel level regularization. To this end, we inject pair of clean and perturbed images during training and minimize the distance between the two resulting embeddings. We study the effects of random noise, low resolution, and mixed resolution due to RoI encoding. We experiment our algorithm for MNIST, CIFAR10 and ImageNet and show improved robustness for perturbed images compared to baseline data augmentation approach.

IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)