An empirical study on the effects of different types of noise in image classification tasks

Gabriel B. Paranhos da Costa, Welinton A. Contato, Tiago S. Nazaré, João do E. S. Batista Neto, Moacir PontiarXiv 2016 · arXiv

Abstract

While computer vision models excel on clean datasets, real-world sensor noise significantly degrades performance. This paper presents a systematic empirical evaluation of how Gaussian, Poisson, and Salt & Pepper noise affect classic image classification pipelines utilizing Local Binary Patterns (LBP) and Histogram of Oriented Gradients (HOG) descriptors. Our findings reveal that noise does more than create a domain gap; it fundamentally compromises the linear separability of image classes. Furthermore, while denoising methods like Non-Local Means (NLM) can mitigate some effects, they frequently remove essential discriminative textures, limiting overall classification accuracy.

Introduction

In controlled laboratory environments, computer vision models achieve high accuracy on "clean" datasets. However, practical applications—ranging from low-light surveillance to satellite imaging—frequently encounter significant sensor noise. Despite the focus on high-level model architectures in recent literature, the "early stages" of the classification pipeline, specifically preprocessing and denoising, are often overlooked. This study addresses whether performance drops in noisy environments stem from a lack of model robustness or if the noise itself renders the underlying data classes inherently harder to distinguish.

Methodology

The research employs a rigorous experimental framework to isolate noise effects within a standard classification pipeline.

The classic image classification pipeline evaluated in this study, highlighting the stages from acquisition to classification.

Feature Extraction and Noise Injection

The study utilizes two primary descriptors: Local Binary Patterns (LBP) for capturing texture and Histogram of Oriented Gradients (HOG) for shape and edge information. Three types of noise were applied at five intensity levels:

  • Gaussian: Simulating electronic circuit noise.
  • Poisson: Simulating photon shot noise.
  • Salt & Pepper: Simulating transmission errors or "dead" pixels.

Denoising and Classification

Two restoration methods were tested: a simple Median Filter and the more advanced Non-Local Means (NLM), which preserves edges through patch similarity. Classification was performed using linear Support Vector Machines (SVMs) in a "one-vs-all" configuration. Separate models were trained for every specific noise type and level to evaluate the performance of "specialized" models.

Datasets

The study utilized two benchmark datasets to ensure diversity in visual content:

  • Corel: 10,800 RGB images across 80 classes.
  • Caltech101-600: A subset of Caltech101 featuring 6 classes (airplanes, bonsai, chandelier, hawksbill, motorbikes, and watch) with 100 examples each.
Representative samples from the datasets used in the empirical study.

Results and Discussion

The Separability Crisis

The most critical finding is that noise reduces the linear separability of classes. Even when models are trained specifically on noisy data, they fail to reach the performance levels of the noise-free baseline. This indicates that noise destroys essential discriminative information that cannot be fully recovered by training on similar noise.

Descriptor Resilience: HOG vs. LBP

HOG descriptors demonstrated greater resilience to noise compared to LBP. LBP's reliance on fine-grained pixel comparisons makes it highly sensitive; even minor noise can flip bits in the LBP code, drastically altering the feature vector.

The Denoising Trade-off

While NLM denoising improved F1-scores for noisy data, it introduced a "performance ceiling." The smoothing effect of denoising filters often removes the fine textures required by descriptors like LBP, illustrating a double-edged sword where restoration can inadvertently strip away useful classification features.

Generalization and Performance Decay

Models trained on clean images generalized poorly to noisy ones. Furthermore, models specialized for one noise type (e.g., Salt & Pepper) showed limited effectiveness when tested against others (e.g., Gaussian). As noise intensity increased, F1-scores dropped significantly; for instance, in the Caltech101-600 dataset using LBP, scores fell from ~0.8 (clean) to below 0.4 under high Gaussian noise.

Conclusion and Future Work

This empirical study highlights the fundamental challenges noise poses to image classification. Future research should expand this analysis to include a wider variety of descriptors, non-linear kernels, and deep learning architectures. Additionally, extending this work to video data could leverage temporal information to mitigate noise effects. Incorporating quality metrics like PSNR and SSIM would also allow for a more precise correlation between visual degradation and classification loss.

Summary updated

BibTeX
@misc{costa2016empirical,
  author = {Gabriel B. Paranhos da Costa and Welinton A. Contato and Tiago S. Nazaré and João do E. S. Batista Neto and Moacir Ponti},
  title = {{An empirical study on the effects of different types of noise in image classification tasks}},
  year = {2016},
  eprint = {1609.02781},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1609.02781},
}
Figure