Classifying with Noise: A First Step into Identifying Occluded Digits

The accurate detection and classification of occluded objects in images present an ongoing challenge in the field of computer vision. In our current work, we tackle a situation where we need to identify the objects that are handwritten digits from the MNIST dataset1. Our proposed approach is decomposed into two steps. In Step 1, a classification model based on Convolutional Neural Networks is designed and implemented to classify digits that are augmented with noise at varying degrees. In Step 2, we propose to introduce a model based on Single Shot Detector and tailor it to our purpose. Once a possible digit is located in an image, it will be then classified by our model from Step 1. While we are still finalizing Step 2 of our proposed approach, the experiment results from the classification model we propose for Step 1 are promising. In this paper, we are presenting the details of this model, including the architecture of the model, hyperparameters, dataset augmentation, etc. We believe that our proposed CNN-based model for classifying noisy digits is also applicable to classifying other objects, such as letters, characters, etc.