Enhancing Image Classification and Explainability with Object Isolation and Background Randomization

In this paper, we address the problem of training a model on an image dataset that contains multiple objects that can introduce noise during the training of an image classification model. We propose a method for separating individual objects from the images and synthesizing these separated objects with a random background dataset to generate a new dataset in which each image contains a single, clearly defined object. We use the Attribution mask Compress-Semantic Input Sampling for Explanation (AC-SISE) method, a perturbation-based explainable artificial intelligence (XAI) model, to analyze the explainability of models trained on the previously generated dataset and the original dataset. The experimental results show that the ResNet50 model does not improve the explainability, but the VGG16 model improves the explainability somewhat.