Enhancing Classification Models With Sophisticated Counterfactual Images

In deep learning, training data, which are mainly from realistic scenarios, often carry certain biases. This causes deep learning models to learn incorrect relationships between features when using these training data. However, because these models have black boxes, these problems cannot be solved effectively. In this paper, we aimed to 1) improve existing processes for generating language-guided counterfactual images and 2) employ counterfactual images to efficiently and directly identify model weaknesses in learning incorrect feature relationships. Furthermore, 3) we combined counterfactual images into datasets to fine-tune the model, thus correcting the model's perception of feature relationships. Through extensive experimentation, we confirmed the improvement in the quality of the generated counterfactual images, which can more effectively enhance the classification ability of various models.