Improving the Robustness of a Convolutional Neural Network with Out-of-Distribution Data Fine-Tuning and Image Preprocessing

Deep convolutional neural networks trained on readily available datasets are often susceptible to decreases in performance when executing tasks on new data from a different domain. Making models generalize well on data in a new domain is the task of domain adaptation. Recently, a simple method, known as Out-of-Distribution Image Detector for Neural Networks (ODIN), was proposed for identifying out-of-distribution (OOD) images in a dataset. This paper proposes fine-tuning an image classifier model using OOD images detected in an ideal training set to improve the model's ability to classify real-life images. This work aims to investigate the effectiveness of such a technique, as well as image preprocessing methods like background removal and image cropping, at increasing the robustness of a ResNet50V2 baseline image classifier in the context of a multi-class classification task. It was observed that fine-tuning with OOD images identified by ODIN consistently increased the model's performance and that a combination of cropping images and fine-tuning with OOD images resulted in the greatest increase in the model's performance.