Robust Deep Stack Auto-encoder Approach for Image Classification: A Novel Fuzzy Attitude

Background and Objectives: Deep learning has been recently integrated with fuzzy logic tolearn efficient models against uncertainty factors. Due to the daily spread of imaging equipment,different types of uncertainty appear in image analysis tasks. In general image analysis applications,uncertainty management is a challenging issue. Furthermore, image analysis issues involveuncertainty in low-level and high-level features, which are not evaluated in related research worksalready. Developing a novel image classification approach to evaluate the uncertainty in severalfeature levels is the main motivation of this study. The main purpose in this paper, a robust deepstack auto-encoder model is proposed for image classification while uncertainty is evaluated inseveral levels of features.Methods: Deep learning-based models have demonstrated outperformance in different imageclassification tasks in recent years. To address uncertainty issues, we employ a fuzzy attitude todeep learning that decrease the effects of uncertainty in image data. A deep stack auto-encodermodel is proposed in this paper in which fuzzy membership degrees are applied to the activationfunction of the neurons in the deep model. More adaptability and especially higher degrees offreedom in fuzzy parameters contribute to better manage the inherent uncertainties in complexdata to a large extent.Results: The experiment is performed on three imbalanced image datasets, including CIFAR-۱۰,Caltexh۱۰۱, and Caltech۲۵۶. The experimental results of the proposed classification method appliedto these datasets demonstrate that the deep stack auto-encoder model with fuzzy attitude canadequately minimize the negative effects of uncertainty in input images.Conclusion: Comprehensive comparisons between the proposed model and some other state-ofthe-art classification methods are performed. The performance results show the outperformanceof the proposed model compared to other recent strategies. Additionally, the evaluation resultsindicate the robustness and efficiency of this model in uncertainty management.