Research on quantification of image classification uncertainty based on conformal prediction

Convolutional neural networks (CNNs) have demonstrated remarkable performance in image classification tasks. However, for real-world decision-making applications, it is crucial to develop reliable methods for quantifying the uncertainty associated with their predictions. This study evaluates two conformal prediction methods, Regularized Adaptive Prediction Sets (RAPS) and Adaptive Prediction Sets (APS), using a ResNet-152 classifier on ImageNet. The focus is on their performance under unperturbed conditions and perturbations such as noise, blur, and occlusion. Experiments analyze varying confidence thresholds and RAPS regularization adaptability. Results show that RAPS outperforms APS in most scenarios but faces challenges in prediction efficiency at high confidence levels. This research provides valuable references for practical applications of uncertainty quantification in image classification tasks, contributing to enhancing the reliability of CNN-based classification systems in complex environments.