TL;DR
This paper tackles the challenge of inconsistent spectral reflectance measurements in integrating-sphere-based multispectral imaging systems. They propose a novel two-stage correction method that first uses a white standard and a reference white patch to address spatial non-uniformity and lighting deviation, followed by a polynomial regression model to further refine accuracy for non-Lambertian samples. This approach significantly improves the consistency and accuracy of spectral and colorimetric measurements.
What Problem Does This Paper Address?
Multispectral imaging systems using integrating spheres are valuable for capturing detailed spectral information across an image. However, these systems face a significant challenge: measurement inconsistency. This arises from two primary sources:
- Spatial Non-Uniformity: Even with a uniform light source, variations in lens transmission and filter characteristics can cause different parts of the same sample to produce different readings.
- Lighting Deviation (Background Effect): The sample itself, through multiple reflections within the integrating sphere, can alter the illumination conditions. This means the same small color patch might yield different measurements depending on the surrounding background. Additionally, fluctuations in the light source contribute to this deviation.
Existing calibration methods often focus on spatial non-uniformity but fail to adequately address lighting deviation, especially for non-Lambertian materials, leading to inaccurate spectral reflectance and color measurements.
Methods
The authors propose a two-stage correction method:
- Stage I: White-Patch Normalization:
- Spatial Non-Uniformity Correction: An image of a large "white standard" (a white board with the same dimensions as the sample holder) is acquired. This allows for the characterization and correction of spatial variations in the camera's response.
- Lighting Deviation Compensation (Lambertian Assumption): A small "reference white patch" is introduced. This patch, placed alongside the sample, acts as a reference for the incident light. By comparing the response of the white patch when imaging the standard versus when imaging the sample, the lighting deviation is compensated for, assuming the sample is Lambertian. This stage effectively introduces a "reference beam" equivalent.
- Stage II: Polynomial Regression Modeling:
- Addressing Non-Lambertian Samples: Recognizing that real-world samples are often non-Lambertian, Stage I's correction is insufficient. This stage employs a polynomial regression model to further reduce measurement inconsistency.
- Model: A second-order polynomial model is used, taking the normalized camera response (after Stage I) and the white-patch ratio as input variables. This model learns the relationship between these inputs and the desired consistent output, effectively learning to remove the background effect.
- Training: The model is trained using a dataset of small gray patches placed on various background samples at different positions. The goal is to ensure that the corrected response of a given patch is consistent regardless of the background.
Datasets
- Training Data: The authors collected training data using a procedure where small gray patches (0.5 cm x 0.5 cm) were placed on larger background samples (10 cm x 8 cm) at various positions. They used 9 small gray patches and 10 background samples. The images of these patches and backgrounds are shown in Figure 3.
- Test Data: The test data also consisted of small color patches placed on various background samples with different patterns (Figure 4). These were distinct from the training samples to validate the generalization capability of the method.
Key Insights and Ideas
- The "Background Effect" is Crucial: The paper highlights that the sample itself can significantly alter the lighting conditions within the integrating sphere, a problem often overlooked or simplified in previous methods.
- Two-Stage Approach for Robustness: Combining a physics-based correction (Stage I) with a data-driven approach (Stage II) provides a powerful solution. Stage I handles the more predictable aspects (spatial uniformity, Lambertian lighting), while Stage II tackles the complex, non-linear effects of non-Lambertian materials.
- White Patch as a Dynamic Reference: The reference white patch acts as a dynamic reference, allowing the system to adapt to lighting changes caused by different samples, effectively mimicking a reference beam without requiring a separate optical path.
- Phenomenological Model for Practicality: While the polynomial model is phenomenological (meaning it describes how things behave rather than why), it proves effective for practical correction, especially when dealing with similar material types (textiles in this case).
Key Results
- Significant Consistency Improvement: The proposed method achieved an 85.3% improvement in camera response consistency (measured by the reduction in standard deviation) compared to the original data.
- Reduced Spectral and Colorimetric Errors:
- For training data, the spectral RMS error dropped from 0.1169 to 0.0149 (Stage II).
- The average color difference error (CIEDE2000 under D65) reduced from 5.1649 units to 0.3679 units (Stage II).
- Similar significant reductions were observed for test data.
- Visual Validation: Figure 5 visually demonstrates the reduction in fluctuations of camera responses across different backgrounds after applying the corrections. Figure 6 shows that spectral reflectance curves become much more consistent after Stage II correction.
- Superior to Existing Methods: Comparison with a Markov model (a relevant state-of-the-art method) showed that the proposed method achieved significantly lower standard deviations for camera responses and lower spectral/colorimetric errors.
Limitations and Future Work
- Generalizability of Polynomial Model: The authors acknowledge that the phenomenological polynomial model might not generalize perfectly to samples with significantly different reflection characteristics than those used in training.
- Material Specificity: The current work focused on textile samples made of similar materials. Extending the method to a wider range of materials with diverse optical properties would be a valuable future direction.
- Cross-Media Calibration: For samples with vastly different optical properties, additional cross-media calibration might be necessary, as mentioned in the paper.
- Further Optimization: While the second-order polynomial was found to be sufficient, exploring higher-order polynomials or alternative regression models could potentially yield further improvements.
Summary updated