Evaluation and comparison of multispectral imaging systems

Multispectral imaging, which extends the number of imaging channels beyond the conventional three, has demonstrated to be beneficial for a wide range of applications. Its ability of acquiring images beyond the visible range and applicability in many different application domains lead to the design and the development of a number of multispectral imaging technologies and systems. Given different systems to choose from, it is important to be able to compare them in a general and in many situations specific to a certain application of interest. In this paper, we evaluate several conventional and recently proposed multispectral imaging systems, both qualitatively and quantitatively. Both spectral and colorimetric accuracies are used as the criteria in the quantitative evaluation. The systems are evaluated and compared for two specific applications: imaging of natural scenes and paintings (cultural heritage), as well as for a general spectral imaging solution. This work provides a framework for the evaluation and comparison of different multispectral imaging systems, which we believe, would be very helpful in identifying the most appropriate technique or system for a given application. Introduction Spectral imaging allows capture of an image of a scene called a spectral image, which represents an environment independent physical property of each surface points of the scene in the form of spectral reflectances. A spectral image can have information beyond the visible range, such as infrared and ultra-violet. There are basically two main types of spectral imaging techniques: hyperspectral imaging and multispectral imaging. Hyperspectral imaging (HSI), which acquires spectral images in a large number of narrow spectral bands, produces high spectral accuracy. However, the acquisition time, complexity and cost of these systems are generally quite high compared to multispectral systems. Multispectral imaging (MSI) on the other hand acquires images in a limited number of relatively wide spectral bands, and the spectral reflectance functions are obtained from the sensor responses using an estimation algorithm. Multispectral imaging provides cheaper and faster solutions compared to hyperspectral imaging with good enough quality for many applications. Multispectral imaging has widespread application domains, such as remote sensing [1], medical imaging [2], biometrics [3], cultural heritage [4, 5] and many others. Many different types of multispectral imaging techniques and systems have been proposed in the literature. In a conventional filter-based imaging system, either a set of traditional optical filters in a filter wheel, or a tunable filter [6, 7] in front of a monochrome camera are employed, and images of a scene are acquired with one each of these filters in a sequence. The use of RGB cameras increases the acquisition speed by three times [8, 9]. Shrestha et al. [10, 11] proposed a single-shot six band multispectral system using a stereo camera (StereoMSI). Another single shot multispectral imaging solution is the filter array (FAMSI) [12–14], which is based on the extension of filter array from 3-channel as in Bayer pattern [15] further, allowing acquisition of more than three band images. Pixel value corresponding to a missing filter at a pixel position is estimated from the neighboring pixels having the filter through interpolation, the process called demosaicing. Another promising technique of multispectral imaging is based on multiplexed LED (Light Emitting Diode) illumination (LEDMSI) [16–19]. In a typical LED illumination based multispectral imaging system, a set of n different types of LEDs are selected, each type of LED is illuminated in a sequence, and a monochrome camera captures an image under the illuminated LED, thus producing an n-band image (n-band RGB-LEDMSI). Shrestha and Hardeberg [20] proposed a three times faster LED illumination based system which uses an RGB camera and optimal combinations of three different types of LEDs that lie on the red, the green and the blue regions bounded by the camera sensitivities. From n-band image acquired with any of the multispectral imaging technique or system, the spectral reflectance image of the scene is obtained using a spectral estimation method. Given different multispectral imaging systems, it is important to evaluate and compare the performance and the quality of these systems. This is useful in identifying a suitable system to be used for a given application. In this paper, we evaluate the three major different types of fast and practical multispectral imaging techniques: Stereo camera, MSFA, and LED illumination based multispectral imaging systems, along with the conventional filter wheel and liquid crystal tunable filter (LCTF) based systems, both qualitatively and quantitatively. We identify important quality attributes and compare the systems based on them. Quantitative evaluation is done based on spectral and colorimetric accuracies from the spectral reflectance image, estimated from the multi-band images acquired by different systems. We present next the multispectral imaging systems and the setups used in our evaluation and comparison. Qualitative and quantitative evaluations of the systems will be presented in the following two sections. The results from the evaluations will be discussed next, and finally we conclude the paper. Multispectral imaging systems and setups used In this section, we briefly describe the three new promising fast and practical multispectral imaging systems: StereoMSI, FAMSI, and LEDMSI and two conventional systems, filter wheel and LCTF based systems. 107 22nd Color and Imaging Conference Final Program and Proceedings and 2nd Congress of the International Academy of Digital Pathology • Filter wheel based MSI (FWMSI) system: In a typical FWMSI system, n number of images of a scene is acquired by a monochrome camera, with each of the filters in the rotating filter wheel placed in front of the camera sensor or the lens of the camera. It acquires an n-band image in n exposures. We use a PixelTeQ’s SpectroCam UV-VIS camera (http:// www.pixelteq.com/product/spectrocam-uv/) with six filters in its filter wheel. The six filters used are the band pass filters with peakwavelength(nm) FWHM of 425 100, 475 100, 505 50, 550 100, 615 100 and 650 100. The spectral transmittances of the six filters are given in Figure 1a. The spectral sensitivity of the SpectroCam camera along with the other two cameras is given in Figure 4. • LCTF based MSI (LCTFMSI) system: In a typical LCTFMSI system, n number of images of a scene is captured by a monochrome camera in n exposures, one each with a narrow (or somewhat wide) band filter activated with an electronically controllable LCTF [6, 7]. We use DVC-16000M monochrome camera and VariSpec-SN50346 LCTF filter. The spectral sensitivity of the DVC-16000M camera is given in Figure 4. The VariSpec-SN50346 LCTF filter has 33 electronically controllable filters. For a fair comparison with the other systems, we use six filters (the same number as in the most of the other systems) selected based on uniform spacing in the visible range (Figure 1b). 400 500 600 700 0 0.2 0.4 0.6 0.8 1