From Human Vision to Neural Networks: Improving Robustness in Degraded Shape Recognition

View at publisher Summarise with VIRAT
Authors: Deepti Pawade, Yash Jadhav, Yash Agashe, Prannoy Banerjee
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Neural networks usually find it difficult to identify objects under degraded conditions like noise, blur, and occlusion—tasks where human vision is still resilient. This paper examines the gap in performance between human vision and deep learning models under degraded shape recognition. We discuss biologically motivated methods like Capsule Networks, Vision Transformers (ViTs), and processes such as reconstruction-guided attention and spatialfrequency processing. These models are tested on CIFAR-10-C with various types of corruption, through metrics like accuracy, mean corruption error (mCE), and robustness score. Experimental outcomes indicate that Capsule Networks and ViTs have significantly higher performance compared to standard CNNs, providing higher resistance against visual degradation. Our results highlight the need for merging cognitive principles and robust testing frameworks to design neural systems that better replicate human-like perception in the real world.

Fields of Study & Subjects

Advanced Neural Network Applications Adversarial Robustness in Machine Learning Machine Learning and Data Classification

Publication details

Journal / source
CRC Press eBooks
Pages
180–184
Publication date
25 March 2026
Language
EN
Document type
Book chapter