1. Introduction
Thermal imaging is being extensively used owing to the ongoing global pandemic of coronavirus disease 2019 (COVID-19). Fever is one of the prominent symptoms of COVID-19, and people infected by the virus are detected efficiently by LWIR thermal cameras (also known as microbolometer cameras). Infrared (IR) thermography converts radiation in the LWIR bands to thermal images (commonly referred to as thermograms), and allows for the detection of temperature variations. Visualization of thermal information benefits the recognition of the surroundings in some cases (e.g., low-light conditions or dark areas) because thermal imaging tends to provide more meaningful information in urban environments with huge thermal variations [
1]. Therefore, LWIR imagers have been extensively used in medical, military, and security applications, As well as commercial applications, such as remote sensing [
2,
3], medical imaging [
4], advanced driver-assistance systems [
5], and face recognition [
6,
7,
8].
Continuing advances in IR imaging have resulted in inexpensive, portable IR cameras, shutterless operations, and both cooled and uncooled, which can provide high-resolution thermograms [
9]. Despite the advances and increasing demand for thermograms, thermal imagers still suffer from certain problems when measuring the thermal distribution within the field of view. Microbolometers are sensitive to the radiation emitted by objects in the scene, and the background radiation reflected by the objects in the scene, and the background radiation reflected by these objects. The relationship between the irradiation collected at the sensor and the temperatures of the imaged materials is nontrivial. For example, the material and surface properties of an object alter its emissivity. Similarly, the reflective properties of an object will vary the amount of background radiation reflected by the object and subsequently collected by the thermal imaging sensor. This variability can lead to errors in the measurement of an object’s temperature.
Additionally, the LWIR imaging sensor suffers from image degradation caused by its intrinsic limitations. First, a typical LWIR camera has a problem of severe noise in the framework without cooling systems. In general, LWIR cameras with cooling systems provide high-quality thermograms; however, applications that use handheld devices require minimized LWIR cameras [
9]. Most LWIR imagers employ uncooled systems because of their size, weight, power, and price. A thermal camera without the cooler is usually referred to as a microbolometer. Second, residual non-uniformity (RNU, also known as fixed pattern noise (FPN)) is another degradation of the LWIR imaging sensor without a mechanical shutter. Typical LWIR cameras employ a shutterless imaging system that allows sensors to operate continually without the need for the mechanical shutter for NUC. The shutterless operation prevents external interruption yielding recalibration, and the changing thermal distributions can be treated based on the temperature measurements of the camera [
10]. However, shutterless operations refer to the absence of a reference LWIR image typically used to cancel the non-uniformity, and RNU are observed owing to the low performance of the NUC [
11].
Various conventional methods have been proposed to reduce the distortions of thermal imaging sensor [
12,
13,
14]. These traditional methods considered adjacent pixels or referred data sequences to reduce the RNU. However, the characteristics or statistics of thermal images have seldom been employed to regulate thermal image processors, even if the statistics of LWIR images have been studied for a decade. Morris et al. [
1] found that the power spectrum of a thermal image is predicted by the generalized Laplace distribution and suggested ranges of the parameter of the Laplace distribution. Goodall et al. [
15] studied distortions unique to LWIR images, and found statistical models for measuring the degree of presetting non-uniformity in thermal images. Moreno-Villamarín et al. [
16] analyzed common image distortions, such as blur, white noise, JPEG compression, and non-uniformity, in fused LWIR and visible light images, and proposed an opinion-aware fused image quality analyzer. Pezoa and Medina [
11] proposed a model for RNU (or non-uniformity) present in IR focal plane arrays (FPAs) in the frequency domain (note that the RNU in the thermal imaging sensor is usually observed as a grid-like pattern and is different from those of complementary metal-oxide semiconductor (CMOS) sensors). In particular, the statistical regularities of LWIR [
16] or fused visible-LWIR images [
15] were derived based on natural scene statistics (NSS) of the visible light images, but the characteristics were not incorporated as prior information to solve optimization problems of thermal imaging.
The thermal image processing algorithm can consider the prior knowledge, which is about the properties of LWIR images. Usually, the image restoration algorithm that considers visible light images attempts to incorporate the characteristics of the images, and the natural image statistics are usually derived from the distribution of gradients. For example, a mixture of the Gaussian approximation [
17] and hyper-Laplacian [
18] priors is derived from the heavy-tailed distribution of gradients in natural scenes and is incorporated as a regularization strategy. Although the relationship between the statistics and qualities of visible light images has been extensively studied, limited work has been conducted on the restoration of thermal images based on considerations of their own statistical characteristics. Our observation of LWIR images indicates that restoration of an LWIR image requires the inherent characteristics of the image. The spectral band of LWIR is characterized by wavelengths in the range of 8 to 14
m and possesses its own statistics in the FPA. The difference between the gradients of visible light and LWIR images is shown in
Figure 1 (note that we analyzed five datasets marked as MORRIS [
1], KRISTO [
19], ADAS [
5], OSU1 [
20], and OSU2 [
21]). LWIR and visible light images have different mean-subtracted contrast normalized (MSCN) coefficients and gradient statistics even if they share a comparable field of view.
The experiments reported herein, present the properties of LWIR images without considering the LWIR imager’s optical or electronical characteristics. Therefore, we have tried to derive an intrinsic property using various datasets. A brief description of datasets follows. Dataset Morris [
1] contains about 450 scenes with 350 outdoor and 100 indoor scenes. Each scene consists of a pair of images (visible light and LWIR image). Dataset Kristo [
19] contains 180 scenes with 80 outdoor and 100 indoor scenes. Each scene also consists of a pair of images (visible light and LWIR image). Dataset ADAS [
5] contains 14,452 annotated thermal images with 10,228 images sampled from short videos and 4224 images from a continuous 144-s video. All videos were taken on the streets and highways in Santa Barbara, California, USA from November to May. Dataset OSU [
20] contains about 500 scenes obtained by surveillance cameras and each scene consists of a single LWIR image. Dataset OSU2 [
21] contains about 10,000 scenes obtained by surveillance cameras and each scene consists of a pair of visible light and LWIR images. Finally, dataset SRIP contains 20 scenes, and each scene consists of a pair of visible light and LWIR images.
This paper is organized as follows: in
Section 2, we derive a property of LWIR images using the aforementioned datasets. In
Section 3, we propose an optimization strategy that incorporates our hypothetical property. In
Section 4, the results of our optimization method are compared with those of the conventional methods. Finally,
Section 5 concludes this study.