A Feature-based Approach for the Recognition of Image Quality Degradation in Automotive Applications

Bauer, Florian · 2023 · Crossref

DOI: 10.1109/iwssip58668.2023.10180245

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical need for reliable image quality assessment in automotive visual perception systems, particularly for automated driving. In-vehicle cameras, especially low-mounted fisheye cameras, are frequently exposed to environmental hazards such as rain, mud, dust, and ice, which cause lens soiling and other degradation effects like blur, glare, noise, and underexposure. These issues significantly impair the performance of perception algorithms. The authors propose a lightweight, feature-based algorithm to detect and classify these degradation types, aiming to enable systems to issue warnings or disable affected functions when image quality is compromised. The proposed method relies on extracting statistical features from single grayscale images, making it applicable to a wide range of automotive cameras, including those without color information. The algorithm first applies five specific filter operations to the input image: local mean subtraction, local contrast calculation, Laplacian gradient computation, Mean Subtracted Contrast Normalized (MSCN) coefficients, and pair-wise products of MSCN coefficients. From the distributions of these five quantities, the algorithm computes the first and second statistical moments (mean and variance) for positive and negative values separately. This process yields a compact feature vector of 20 scalar elements. These features are then fed into a Support Vector Machine (SVM) classifier with a Radial Basis Function kernel, chosen for its efficiency and robustness with small datasets. Experiments were conducted using four distinct datasets comprising over 19,000 images, including the public WoodScape dataset and proprietary data from test vehicles. For binary classification of clean versus soiled images, the algorithm achieved accuracies between 95.51% and 99.75% across the different datasets. In a multi-class classification task involving six degradation types (clean, soiled, blur, glare, noise, underexposure), the algorithm achieved an overall accuracy of 96.52%. When compared to a Convolutional Neural Network (CNN) approach on the WoodScape dataset, the feature-based method demonstrated superior generalization capabilities. Specifically, when trained on front and rear camera images and tested on all camera positions, the feature-based approach achieved 91.76% accuracy, outperforming the CNN’s 83.97% accuracy, despite using significantly less training data. The study concludes that the feature-based approach offers a highly efficient and effective solution for detecting image quality degradation in automotive applications. Its primary advantages are the ability to achieve high accuracy with small training sets and low computational complexity, making it suitable for real-time implementation. Furthermore, the method’s strong generalization across different camera perspectives suggests it is robust for practical deployment in automated driving systems, where detecting and mitigating the impact of degraded sensor data is essential for safety.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• disability glare