Understanding how noise affects the accuracy of CNN image classification

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Convolutional neural networks (CNNs) remain the standard approach to image classification, but their accuracy can drop sharply when the input images are degraded by noise introduced during acquisition or transmission.This paper describes a single experimental protocol for measuring that drop under three noise types: Gaussian, impulse (also called salt-and-pepper), and speckle noise.Each type is applied at several increasing intensities so that the effect of noise can be followed across severity levels.CIFAR-10 serves as the benchmark, and a CNN trained on clean images is the reference model.All noisy test sets are built from the same original images using predefined severity schedules, which keeps the architecture, labels, preprocessing, and evaluation settings identical across conditions.The aim is to separate the effect of noise from the rest of the pipeline.Evaluation uses top-1 accuracy, error rate, macro-precision, macro-recall, macro-F1, accuracy degradation, and class-wise confusion, which together show not only whether the model fails but also how it fails.A second condition, reported separately, adds noise augmentation during training to test whether earlier exposure makes the model more robust.The manuscript covers the research motivation, related work, experimental protocol, noise definitions, implementation controls, analysis plan, limitations, and reproducibility requirements, and it follows IEEE style.Because no execution logs or measured predictions were available for this revision, no numerical results are reported; the protocol only specifies how empirical results should be recorded and interpreted.