CoChat
Sign Up
Cited by 8Image and Video Quality AssessmentAdvanced Image Processing TechniquesImage Enhancement Techniques

No-Reference Image Quality Assessment with Moving Spectrum and Laplacian Filter for Autonomous Driving Environment

Open access
Nam, Woongchan, Youn, Taehyun, Ha, Chunghun·2025·Vehicles
DOI 10.3390/vehicles7010008
Open this paper in CoChat — free accountOpen PDF

Abstract

The increasing integration of autonomous driving systems into modern vehicles heightens the significance of Image Quality Assessment (IQA), as it pertains directly to vehicular safety. In this context, the development of metrics that can emulate the Human Visual System (HVS) in assessing image quality assumes critical importance. Given that blur is often the primary aberration in images captured by aging or deteriorating camera sensors, this study introduces a No-Reference (NR) IQA model termed BREMOLA (Blind/Referenceless Model via Moving Spectrum and Laplacian Filter). This model is designed to sensitively respond to varying degrees of blur in images. BREMOLA employs the Fourier transform to quantify the decline in image sharpness associated with increased blur. Subsequently, deviations in the Fourier spectrum arising from factors such as nighttime lighting or the presence of various objects are normalized using the Laplacian filter. Experimental application of the BREMOLA model demonstrates its capability to differentiate between images processed with a 3 × 3 average filter and their unprocessed counterparts. Additionally, the model effectively mitigates the variance introduced in the Fourier spectrum due to variables like nighttime conditions, object count, and environmental factors. Thus, BREMOLA presents a robust approach to IQA in the specific context of autonomous driving systems.

AI snapshot

Written from the paper’s full text.

TL;DR

BREMOLA is a no-reference image quality assessment metric for autonomous driving cameras that combines the central area of a shifted Fourier magnitude spectrum with Laplacian-filter edge complexity to detect blur from camera degradation. On 45 driving frames with simulated average-filter blur, it reduced metric variance and separated normal from degraded images better than PSNR, SSIM, GMSD, and BRISQUE.

Key findings

  • BREMOLA computes quality as the central white area \(A\) of the moving spectrum divided by the square root of Laplacian edge complexity \(B\); Laplacian was selected because it had the highest SROCC with the moving-spectrum central area at 0.9236, versus Canny 0.8879, Sobel 0.7113, Scharr 0.4581, Prewitt 0.3892, and Roberts 0.2912.
  • Using 45 driving frames processed at 20 blur levels from \(1\times1\) to \(20\times20\), yielding 900 samples, BREMOLA reduced the range of metric values and reduced overlap between blur levels; before BREMOLA, the \(1\times1\) and \(3\times3\) blur levels overlapped by half.
  • In comparisons, PSNR, SSIM, and GMSD showed large metric differences and inversions at blur levels \(1\times1\) to \(5\times5\), while BRISQUE had no inversion but too-large inter-image differences; BREMOLA had small inter-image differences and distinguished image quality degradation.
  • In a Seoul driving-footage monitoring simulation sampled once per second with blur increased every 30 s (\(1\times1\) for 0–30 s, \(2\times2\) for 30–60 s, \(3\times3\) for 60–90 s, \(4\times4\) after 90 s), BREMOLA values all dropped below the failure threshold after 60 s, whereas without BREMOLA values exceeded the threshold at the failure point and caused false alarms.

Methods

  • Dataset: 15 frames were randomly extracted from each of three driving videos from Dubai, Los Angeles, and San Francisco, totaling 45 distinct frames; environmental factors varied included time of day, number of people and cars, and presence of buildings. All images had resolution \(1920\times1080\).
  • Blur simulation: each frame was processed with average filters at 20 levels from \(1\times1\) to \(20\times20\), producing \(45\times20=900\) samples; \(1\times1\) is equivalent to the original image.
  • BREMOLA pipeline: convert to grayscale, apply FFT, log-transform, shift the origin to obtain the moving spectrum, threshold pixel values above 100 to count central area \(A\), convolve with a Laplacian filter using zero padding, sum edge pixel values for complexity \(B\), and compute the metric as \(A/\sqrt{B}\). A scaling constant of 14 was used, and experiments were written in Python.

Limitations

  • Stated: the method only applies to blur among various distortion phenomena and must be verified for other application fields and image distortion phenomena; it also needs validation on more diverse driving images, including snow, rain, thunder, and unpaved roads, and must consider the driving decision-making system.
  • Stated: the study excludes physical camera damage such as image sensor destruction or circuit failure, and treats noise, glare, and motion artifacts as temporary environmental distortions outside scope; it assumes aging/degradation mainly manifests as blur.
  • Evident from the text: the experimental dataset is small, using 45 frames from three videos plus a Seoul monitoring example, and blur is simulated with average filters rather than real camera aging; the \(3\times3\) normal versus \(4\times4\) faulty threshold is based on visual inspection and a scree-plot-like heuristic, with no MOS or direct human visual system validation reported.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/vehicles7010008

Cites

24 of 24 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

  • Findingown data
    § 5.3. Comparison with Other Metrics · p. 15

    For BREMOLA, the difference in metric values between images is small.

    “BREMOLA can be used as a metric because the difference in metric values between images is small”
    Quote in the paperSubstantial claim
  • Findingown data
    § 5.2. Experimental Results · p. 14

    The metric values of the 1 × 1 blur level and the 3 × 3 blur level overlap by half.

    “the metric values of the 1 × 1 blur level and 3 × 3 blur level overlap by half”
    Quote in the paperSubstantial claim
  • Findingreasoning
    § No-Reference Image Quality Assessment with Moving Spectrum and Laplacian Filter for Autonomous Driving Environment · p. 1

    BREMOLA effectively mitigates variance introduced in the Fourier spectrum due to variables like nighttime conditions, object count, and environmental factors.

    “Additionally, the model effectively mitigates the variance introduced in the Fourier spectrum due to variables like nighttime conditions, object count, and environmental factors.”
    Quote in the paperSubstantial claim
  • Findingown data
    § 3.3. Limitations of NR-IQA Metrics Using Shifted Spectrum · p. 8

    The central area metric has large variance across driving images.

    “First, metrics that use the area in the center of the traveling spectrum are unreliable due to the large variance in size across driving images.”
    Quote in the paperSubstantial claim
  • Findingown data
    § 4.2. BREMOLA (Blind/Referenceless Model via Moving Spectrum and Laplacian Filter) · p. 12

    The right Laplacian kernel has a higher SROCC value than the left kernel.

    “In this paper, we apply the right kernel, which has a higher SROCC value among the two kernels”
    Quote in the paperNot a substantial claim
  • May combine two claims. 18
  • Classified as probably not a substantial claim. 6
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingown data
    § 5.4. Utilizing BREMOLA Metrics · p. 16

    Without BREMOLA, the values exceed the threshold even at the failure point, leading to an increase in false alarms.

    “without BREMOLA, the values exceed the threshold even at the failure point, leading to an increase in false alarms.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The lowest metric value up to the failure point at 60 s (3 × 3 filter) represents the failure threshold.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingreasoning
    § 3.2. Calculating the Magnitude Spectrum and Movement Spectrum · p. 6

    Applying a log transformation to the magnitude spectrum compresses its dynamic range and enhances the visibility of frequency components.

    “To compensate for this, a log transformation can be applied as a normalization as shown in Equation (6). In Equation (6), Q ( u , v ) represents the logarithmically scaled magnitude of Y ( u , v ) , used to compress the dynamic range and enhance the visibility of the frequency components:”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 3.3. Limitations of NR-IQA Metrics Using Shifted Spectrum · p. 9

    The central area is large when there are many elements such as luminous objects, buildings, and cars, and small when the driving environment is simple.

    “As shown in Figure 8, the central area is large when there are many elements such as luminous objects, buildings, and cars, and the central area is small when the driving environment is simple as shown in Figure 9.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingreasoning
    § 3.2. Calculating the Magnitude Spectrum and Movement Spectrum · p. 6

    The shifted spectrum relocates the origin of the magnitude spectrum to emphasize specific frequency components and allows intuitive observation of high-frequency components by gathering them to the center.

    “This process aims to relocate the central point in the frequency domain to emphasize specific frequency components or to facilitate their analysis. It can be seen that a center shift like Figure 4 allows us to intuitively observe the magnitude of the high-frequency components by gathering the high-frequency components that exist in the four corners to the center.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingreasoning
    § 3.2. Calculating the Magnitude Spectrum and Movement Spectrum · p. 6

    The magnitude spectrum of an image appears featureless black because the low-frequency DC component has very large values while the high-frequency AC component has relatively low values.

    “The magnitude spectrum of the image appears to be a featureless black color because the low-frequency component of the image, the direct current (DC), has very large values, while the high-frequency component of the image, the alternating current (AC), has relatively low values.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 5.4. Utilizing BREMOLA Metrics · p. 16

    With BREMOLA, all values drop below the failure threshold after 60 s, correctly detecting the failure.

    “with BREMOLA, all values drop below the threshold after 60 s, correctly detecting the failure.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The lowest metric value up to the failure point at 60 s (3 × 3 filter) represents the failure threshold.

  • The quote only partly states this claim.

    Findingown data
    § 4. BREMOLA (Blind/Referenceless Model via Moving Spectrum and Laplacian Filter) · p. 11

    Among the compared high-pass filters, the Laplacian HPF has the largest Spearman Rank-Order Correlation Coefficient (0.9236) between edge-detected image complexity and the central area of the moving spectrum.

    “so in this paper, the Laplacian HPF with the largest SROCC is selected for the experiment.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 3.3. Limitations of NR-IQA Metrics Using Shifted Spectrum · p. 8

    Across 45 driving images, the central area of the shifted spectrum maintains a monotonic decreasing function with increasing blur intensity, with no reversal.

    “Regardless of the driving image, the central area maintains the form of a monotonic decreasing function according to the degree of blur, so there is no reversal of the indicator, so it can be used as an indicator, but there are some limitations in applying this method directly to the camera sensor of an autonomous vehicle.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 3.3. Limitations of NR-IQA Metrics Using Shifted Spectrum · p. 9

    Some driving images with a 3×3 to 5×5 average filter applied have central area metric values greater than the minimum value of the original image, making it impossible to determine failure solely based on the metric values.

    “However, as shown in Figure 7, some driving images with a 3 × 3 to 5 × 5 average filter applied have metric values greater than the minimum value of the original image, making it impossible to determine failure solely based on the metric values.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 5.2. Experimental Results · p. 14

    After applying BREMOLA, the range of image quality metric values decreases and the overlap of metric values between blur levels decreases.

    “after applying BREMOLA, we can see that not only does the range of image quality metric values decrease but also the overlap of the metric values between blur levels decreases”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 5.2. Experimental Results · p. 14

    Before applying BREMOLA, image quality is difficult to distinguish because the range of image quality metric values is large and the overlapping area of metric values is large even as the blur level increases.

    “Before applying BREMOLA, it is difficult to distinguish image quality because the range of image quality metric values is large, and the overlapping area of metric values is large even as the blur level increases.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 3.4. Criteria for Determining Camera Deterioration · p. 10

    The average value of the metric across all driving images decreases sharply before the 3 × 3 averaging filter and then decreases slowly after the 3 × 3 averaging filter.

    “The average value of the metric decreases sharply before the 3 × 3 averaging filter and then slowly decreases after the 3 × 3 averaging filter.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationnone given
    § 6. Discussion · p. 17

    The proposed evaluation method is limited in that it applies only blurring among the various distortion phenomena and can be applied in an autonomous driving environment.

    “The evaluation method proposed in this paper has the limitation that it only applies blurring among various distortion phenomena in the existing image quality evaluation area and can be applied in an autonomous driving environment.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § 3.3. Limitations of NR-IQA Metrics Using Shifted Spectrum · p. 8

    The central area metric can be used as an indicator because there is no reversal.

    “Regardless of the driving image, the central area maintains the form of a monotonic decreasing function according to the degree of blur, so there is no reversal of the indicator, so it can be used as an indicator, but there are some limitations in applying this method directly to the camera sensor of an autonomous vehicle.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning
    § 3.2. Calculating the Magnitude Spectrum and Movement Spectrum · p. 7

    The distribution of low-frequency and high-frequency components plays a crucial role in evaluating overall image quality and analyzing the impact of distortions or noise.

    “The distribution of these low-frequency and high-frequency components plays a crucial role in evaluating the overall quality of the image and analyzing the impact of various distortions or noise.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning
    § 1. Introduction · p. 2

    The main information for object recognition is the image captured by the camera, and degradation eventually manifests as a reduction in image clarity.

    “This assumption is justified because the main information that an autonomous system can acquire for object recognition in real time is the image captured by the camera, and degradation eventually manifests itself as a reduction in image clarity.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § 1. Introduction · p. 2

    In autonomous driving, it is impossible to secure original images and practically impossible to build a uniform image dataset for real-time road environments.

    “In autonomous driving systems, the environment in which camera sensors acquire images is a highly variable environment in which buildings, roads, cars, pedestrians, traffic lights, etc., change in real time, so it is impossible to secure original images, and it is practically impossible to build a uniform image dataset for real-time road environments under these conditions.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § 3.4. Criteria for Determining Camera Deterioration · p. 10

    The metric is sensitive to changes in blur before the 3 × 3 median filter and becomes less sensitive afterward.

    “This suggests that the metric 3 × 3 is sensitive to changes in blur before the median filter, and becomes less sensitive afterward.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationown data
    § 3.4. Criteria for Determining Camera Deterioration · p. 10

    In this study, images processed with a 3 × 3 average filter or smaller are considered normal, while those processed with a 4 × 4 average filter or larger are classified as faulty.

    “Based on this, in this study, images processed with a 3 × 3 average filter or smaller are considered normal, while those processed with a 4 × 4 average filter or larger are classified as faulty images for further research.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning
    § 6. Discussion · p. 17

    In autonomous driving, the image acquired by the camera is the original image itself, so FR-IQA and RR-IQA methods cannot be used.

    “In autonomous driving, the image acquired by the camera is the original image itself, so these IQA methods cannot be used.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Most existing image quality assessment methods are FR-IQA or RR-IQA, which require an original reference image to compare against.

  • May combine two claims.

    Interpretationown data
    § 3.4. Criteria for Determining Camera Deterioration · p. 10

    Because the pattern of the metric-versus-blur graph is similar to a scree plot, the same principle can be applied to set the threshold where image quality degrades rapidly.

    “Since the pattern of the graph is similar to a scree plot, which determines the number of factors in a principal component analysis, it is possible to apply the same principle to the threshold where the image quality degrades rapidly.”
    Quote in the paperSubstantial claim
21 of 21

  • The authors state this as an assumption.

    Blur is the primary aberration/degradation in images captured by aging or deteriorating camera sensors, and the study limits its scope to blur.

    “Given that blur is often the primary aberration in images captured by aging or deteriorating camera sensors, this study introduces a No-Reference (NR) IQA model termed BREMOLA (Blind/Referenceless Model via Moving Spectrum and Laplacian Filter).”
    “Furthermore, this degradation is mainly limited to blur. This assumption is justified because the main information that an autonomous system can acquire for object recognition in real time is the image captured by the camera, and degradation eventually manifests itself as a reduction in image clarity.”

  • The authors state this as an assumption.

    The magnitude spectrum obtained by 2D-DFT is a periodic function with origin symmetry.

    “Since 2D-DFT is only possible for periodic functions with origin symmetry, 2D-DFT is based on the assumption that the magnitude spectrum is also a periodic function with origin symmetry:”

  • The authors state this as an assumption.

    The complexity of the image (borderline elements) affects the distribution of the central area in the shifted spectrum.

    “It can be assumed that the complexity of the image, i.e., the borderline elements in the image, affects the distribution.”

  • The paper relies on this without support.

    Based on images alone, it is impossible to distinguish whether an image is acquired from a normal camera or a degraded camera.

    “Based on images alone, it is impossible to distinguish whether an image is acquired from a normal camera or a degraded camera.”

  • The paper relies on this without support.

    Failure, malfunction, or deterioration of camera sensors can produce distorted images that can lead to serious traffic accidents if the vehicle does not recognize lanes or objects correctly.

    “The failure, malfunction, or deterioration of camera sensors can produce distorted images, which can lead to serious traffic accidents if the vehicle does not recognize lanes or objects correctly.”

  • The paper relies on this without support.

    In autonomous driving, the driving environment changes in real time, making it very difficult to control the complexity of the driving environment to derive a reliable indicator with low variance.

    “In autonomous driving, where the driving environment changes in real time, it is very difficult to control the complexity of the driving environment to derive a reliable indicator with low variance.”

  • The paper relies on this without support.

    For a metric to serve as an indicator for determining camera failure or degradation, the metric value of the original image with a 1×1 average filter applied in a normal state must be greater than the metric value with an N×N average filter applied in an abnormal state.

    “For example, the metric value of the original image with a 1 × 1 average filter applied in a normal state must be greater than the metric value with an N × N average filter applied in an abnormal state to serve as an indicator for determining camera failure or degradation.”

  • The paper relies on this without support.

    Most existing image quality assessment methods are FR-IQA or RR-IQA, which require an original reference image to compare against.

    “Most of the existing IQAs are FR-IQA or RR-IQA, which require an original image to compare against.”

  • The authors state this as an assumption.

    The only failure phenomenon caused by aging or degradation of the camera sensor due to prolonged use is blurring, and the tendency of the metric value to decrease with blur intensity can be quantified through the Fourier transform and the resulting shift spectrum.

    “In this paper, we assume that the only failure phenomenon caused by aging or degradation of the camera sensor due to prolonged use is blurring, and that the tendency of the value to decrease with the intensity of the blurring can be quantified through the Fourier transform and the resulting shift spectrum.”

  • The authors state this as an assumption.

    The lowest metric value up to the failure point at 60 s (3 × 3 filter) represents the failure threshold.

    “When assuming that the lowest value up to the failure point at 60 s (3 × 3 filter) represents the failure threshold”

  • The paper relies on this without support.

    Boundaries have the greatest impact on the area in the center of the shift spectrum.

    “Since boundaries have the greatest impact on the area in the center of the shift spectrum, this paper aims to quantify image complexity using edge detection.”
11 of 11

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • The 3 × 3 blur level is classified as both normal and faulty

    Major

    The stated classification treats images processed with a 3 × 3 average filter as normal. The monitoring experiment instead calls the 3 × 3 interval beginning at 60 s the failure point, changing the criterion used to claim successful failure detection.

  • The dataset's blur filter is called both median and mean

    Minor

    The dataset description says a median filter is applied to each image, but immediately describes the applied blur as a mean filter that averages neighboring pixels. These are different filtering operations.

2 of 2

Cites

Flagged first, then by citations

1 of the 25 references couldn’t be matched to a known paper.

Research works printed but not linked: 1

  1. Mlsna, P.A., and Rodríguez, J.J. (2009). The Essential Guide to Image Processing, Academic Press.

Cited by

Sorted by citations

Related

Authors
Nam, W., Youn, T., & Ha, C.
Affiliations
  • Korea University (KR)
  • Research Institute, NURILAB Co., Ltd., Seoul 07228, Republic of Korea
  • Hongik University (KR)
Corresponding author
Ha, Chunghun
ORCID
Venue
Vehicles
Publisher
MDPI AG
Published
2025-01-21
Volume
7
Issue
1
Pages
8
ISSN
2624-8921 (ISSN-L)
DOI
10.3390/vehicles7010008
OpenAlex
W4406671409
Cited by
8
References
25
Citations by year
2026: 4 · 2025: 4
FWCI
3.15
Type
journal article
Language
en
Open access
Licence
Locations
  • Vehicles (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
National Research Foundation of Korea (NRF-2020R1F1A1066272)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image and Video Quality Assessment
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Image Processing Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image Enhancement Techniques
Keywords
no-reference image quality assessment, image blur detection, Fourier spectrum, Laplacian filter, human visual system, autonomous driving, night imagery, image sharpness

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review