This PDF file contains the front matter associated with SPIE Proceedings Volume 12785, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

With the rapid development of the transportation industry, railway transportation plays a crucial role. Manual inspection methods are time-consuming, labor-intensive, and highly subjective. Therefore, there is an urgent need for a more efficient and accurate flaw detection method. This system is a portable rail flaw detection device based on machine vision, with YOLOv5 as its core deep learning algorithm. The system captures surface images of the rail through a camera and transmits them in real-time to the host computer for analysis. Leveraging the powerful real-time object detection capability of YOLOv5s, the system can accurately identify and locate various types of rail surface damages, such as cracks, fractures, and wear. Compared to traditional manual inspection, this system is more efficient and greatly improves the accuracy and efficiency of rail flaw detection. It has a smaller size and is convenient to carry, making it suitable for working in various environments and conditions, greatly enhancing the practicality and flexibility of the device.

The metamer mismatch volume has important applications in color correction, camera design, and light source design. The method based on spherical sampling to calculate the metamer mismatch volume has a long computation time, a large number of duplicated boundary points, too few effective vertices, and the dimension of its metamer set will appear lower than the theoretical dimension. In this paper, we propose a high-dimensional spherical sampling method that samples the metamer set directly, and find all boundary points by selecting direction vectors and polarizing all directions. The experimental results show that our method improves the above problems, the computational speed is faster, the computational results are close, the repetition rate of boundary points is greatly reduced, and the actual dimensionality of the corresponding metamer set is consistent with the theoretical dimensionality.

The choice of light source affects the accuracy of the spectral sensitivity estimation. In this paper, we propose to estimate the spectral sensitivity function of digital camera using spectrally tunable LED light sources. The spectral power distribution of the LED light source is determined by a combination of multiple LEDs and their weight coefficients. The method of tuning the weight coefficients of the LEDs includes Monte Carlo method and particle swarm optimization algorithm, so that the LED light source with the smallest estimation error is defined as the optimal light source. Experimental results show that the particle swarm algorithm gives the best estimation results. The relative error of estimation using LED light sources is significantly reduced when compared with the results when using a single light source for estimation (e.g., D65 light source).

The spectral reflectance of multispectral images can provide more valuable information about object characteristics. In order to improve the utilization of the spectrum, the reflectance reconstruction requires the same system calibration and illumination of the image acquisition. Therefore, Khan proposed the concept of multispectral constancy, which is to transform the multispectral image data into a standard representation through spectral adaptive transformation. Khan used the linear mapping method to solve SAT to convert the multispectral image data obtained under unknown illumination into the image data under standard light source. In order to further improve the spectral utilization rate and expand the application range of multispectral cameras, an algorithm to improve multispectral constancy based on chromatic aberration index is proposed in this paper. The algorithm uses chromatic aberration as the objective function to solve the spectral adaptive transformation. In this paper, ten light sources are used as unknown light sources, SFU and X-rite are used as training and testing datasets, and multispectral camera channels are simulated by Equi-Gaussian and Equi-Energy filters with different number of channels to train and test 5, 6, 8, and 10 channels of data. In this paper, the color difference under different light sources is used as the evaluation index to test the performance of the proposed algorithm, and compared with the Khan method for calculating SAT multispectral constancy. The experimental results show that the spectral constancy algorithm based on color difference can perform better, and expand the application of different kinds of unknown light sources in multispectral constancy.

To improve the color reproduction and realism of digital cameras and to promote the development of computer vision. Camera colorimetry is conditioned on the spectral sensitivity response of the camera being a linear transformation of the color matching function of the human visual system. Previous methods have proposed placing well-designed filters in front of the camera to produce a sensitivity that well matches the Luther condition. In this paper, we optimize the latest matching illumination method (by using a spectral-tunable illumination system to modulate the spectrum of certain light sources), improve the method of designing filters and add new constraints. Experiments demonstrate that the matching illumination method using new objective functions give a 5% improvement over the original method, and the optimization of the filter using a gradient ascent algorithm and a genetic algorithm gives a 10% improvement in chromaticity over the original method. The method of limiting the average transmittance also has a 10% improvement over the previous one. As a result, these methods can make the imaging of digital cameras more accurate and realistic.

The Spectral Reconstruction (SR) algorithm attempts to recover hyperspectral information from RGB camera responses. This estimation problem is usually formulated as a least squares regression, and because the data is noisy, Tikhonov regularization is reconsidered. The degree of regularization is controlled by a single penalty parameter. This paper improves the traditional cross validation experiment method for the optimization of this parameter. In addition, this article also proposes an improved SR model. Unlike common SR models, our method divides the processed RGB space into different numbers of neighborhoods and determines the center point of each neighborhood. Finally, the adjacent RGB data and spectral data of each center point are used as input and output data for the Radial Basis Function Network (RBFN) model to train the SR regression of each RGB neighborhood. This article selects MRAE and RMSE to evaluate the performance of the SR algorithm. Through comparison with different SR models, the methods proposed in this article have significant performance improvements.