The enhancement of low-light images are found to be useful in many application fields, including satellite remote sensing, security surveillance, autonomous driving, and rail transport. Most of the approaches on image enhancement are based on convolutional neural networks and machine learning [8-11]. Zhou, J. et al. built a double-teacher knowledge distilling network for enhancing underwater images. In their framework, a dynamic teaching strategy was applied, which gained improvement of the network’s ability to handle complicated underwater environment [8]. In one work, it is found that underwater image enhancement can be done via building a dual-branch attention mechanism, where the network takes cross-view inputs and utilizes the feature alignment module to explore degradation from different views [9]. In another work, it can be seen that the dual-color space and dense multiscale attention blocks can be utilized to enhancing images, where a color-guided map is introduced to guide the network to process color- degradation information. This can enormously improve color restoration [10]. In some cases, the images are degraded that a lot of structural and statistical information are destroyed. This can lead to the drift of the feature representation in a global scale. To solve these issues, the researchers established estimation about the degree of feature drifts using a statistical mechanism [11]. A variational model was further developed to achieve much correction of the color. In addition, a histogram-equalization was performed to improve the image contrast [11] to get the ultimate enhancement of the images.

Another aspect in this research field is the application of Retinex theory [12-16], where the low-light images are decomposed into several layers. Zhou, J. et al. combined underwater noise, texture, and gradients into a variational Retinex model to reduce noise amplification [12]. Li, X. et al. developed a method called as Deep Parametric Retinex Decomposition, where three modules of parametric Retinex decomposition, enhancement, and refinement are combined [13]. This new Retinex model can greatly avoid color deviation. Liu, W. et al. employed termed Reflectance-Correction Retinex for enhancing thermal images acquired from carbon fiber reinforced polymers [14]. Li, C. et al. constructed a fractional structure and texture aware model, which is integrated in the Retinex model to get the image enhancement [15]. Ma, T. et al. used a multi-scale feature extraction module using the framework of the Retinex theory, which showed the capability of extracting image details and structural information [16]. All these works show the vivid application of the Retinex theory in the image enhancement. A database may be required for using these approaches. It brings difficulty for those who are not professional at artificial intelligence and who have limited resources in databases. Simple and quick approaches are constantly in need.

Near infrared imaging is not easy to be done in the research lab, given that the near infrared camera contains a lot of thermal noise [17-20]. Especially, the intensity of near infrared light can be very low in the imaging [21]. This generated those images with very low intensity and dilute features. A convenient method is required for processing those images acquired from the near infrared imaging. Herein, we have proposed three frameworks by combing a wavelet and a set of simple functions. Those images acquired from blue- fluorescence imaging, low-light conditions, and near infrared camera were enhanced. Furthermore, these frameworks were used for image fusion, especially involving low light images as the input images. Thirdly, their application for extracting the profiles in underwater imaging was explored. In Sec. 2, how the frameworks were constructed was presented in details. In Sec. 3, several kinds of research outcome for these filters were shown. The limitations of our current study and future works were discussed. In Sec. 4, we summarized our finding. Our frameworks are presenting approaches simple and effective, which are envisioned to play an important role in advanced imaging and fluorescence microscopes.

2. Materials and Methods

The blue-fluorescence images were taken by a lab-built facility. A UV-laser with emission wavelength of 350-355 nm was used to irradiate the samples from the side. A blue-light filter was put above the sample which would reduce the UV light irradiate. A camera was connected with the blue-light filter and collect the blue fluorescence image. The low-light images were taken by a cell phone (iPhone Xs Max).

The images acquired from a near infrared camera were obtained from a lab-built setup. A near-infrared laser with emission wavelength of 1064 nm was used to irradiate the samples. A near infrared camera was put above the sample to capture the image of the sample.

Three filters were developed using simple mathematical functions. Their frameworks are shown in Algorithms 1-3. In these frameworks, a lifting wavelet was firstly used to reduce some of the inherent noise in the images. Moreover, several functions were tested for the enhancement of the main features of the images. If one specific set of functions were found to be effective, a continuous operation would be followed using these functions. Ultimately, the filtered values could be used for constructing a new output image.

Figure 6: The sample blue fluorescence images are processed by the Log filter. (a) Tube1 processed. (b) Tube2 processed. (c) Tube3 processed. (d) Pipette processed. (e) Bottle processed.

Table 1: PSNR values of the input images processed by the three filters

Figure 10: Sample images acquired from low-light condition were processed by the Log filter. (a) Weak37 processed. (b)Weak42 processed. (c)Weak43 processed. (d) Weak44 processed. (e) Weak45 processed.

Figure 14: The images acquired from a near infrared camera were processed by the Log filter. (a) Nir4 processed. (b) Nir5 processed. (c) Nir6 processed. (d) Nir7 processed. (e) Nir9 processed.

Figure 25: The underwater images processed by the Sobel operator. (a) Water13. (b) Water14. (c) Water15. (d) Water17. (e)Water19.

3.5. Comparing the performance of a matched filter

We used a matched filter to compare the performance [22]. Figs.17a-17e show the output of the blue fluorescence images processed by the matched filter. Figs.18a-18e show the output of the low-light images processed by the matched filter. Figs.19a-19e show the output of the near-infrared images processed by the matched filter. Table 4 shows the corresponding PNSR values associated with those images processed by the matched filter. The average values of the PSNR are 3.2111, 4.7248, and 0.5284 separately for the blue fluorescence images, the low-light images, and the near-infrared images. Clearly, the images acquired via the matched filter do not look as good as those obtained through our filters (see Figs.17-19).

3.6 Application in image fusion

The images obtained from different imaging techniques, including visible and near infrared imaging, can provide a vivid and comprehensive understanding of the features of the samples being studied. Therefore, the technology of the image fusion has been widely used in industry and medical care [23-26]. Given that our filters are able to enhance the low light images, we used them for the image fusion. Here, our filters were used to extract the low light images firstly. Then, a wavelet transform was applied to combine the frequency of the two input images together. Finally, an inverse of the wavelet transform was used to get the final fused image. Figs.20a-20e showed the fusing impact. We used one low-light image as the input image. Our three filters can capture those key features in the low-light image and fused those features in the output image. They showed the effectiveness in the image fusion involved with low-light images.

3.7 Extracting profiles of the image acquired from underwater condition

Getting images with clear features are essential prerequisite for seabed resources exploration and marine environment monitoring, which makes the profiles extraction from the underwater images is very important [27-31]. The complexity of water and lighting environment makes underwater images may go through color distortion and contrast loss. Therefore, new methods or frameworks should be continuing to develop for this purpose. We used our filters to process the underwater images. Figs.21a-21e are 5 input underwater images. Figs.22-24 are the processed images via our three filters. It can be seen that they showed the effectiveness for extracting the key profiles of the underwater images. We used a Sobel operator for comparing (see Figs.25a- 25e). It can be noted that the Sobel operator is not good enough to acquire the main profile for the underwater images. For example, the profiles of the rock, the bottle, and the sweet potato were not shown in the output images processed by the Sobel operation.

4. Discussion

One of the limitation in our framework is that we have to adjust the parameter and the constants in our filters in order to have a best output. For us, it is not a difficult job given that we are proficient with the operation of our framework. This may be a challenge for those who are not familiar with our framework. Future research would be focused on making GUI that can integrate all the parameter and the constants.

It has been shown the images acquired with the assistance of the nanoparticles and special small molecules can show good image contrast and high penetration depth [32-38]. This may be due to the existence of the nanoparticles that can change light scattering and absorption. This would be our future endeavor.

Our work can raise the general interest of the readers in the field of electrical engineering given that the processing of the fluorescence images, the underwater images, the low-light images, and the near infrared images are required in many instruments in this field. Our work can provide general design of filters for the image enhancement. It can provide a worthwhile about setting up a useful filter for processing the low-light images, the underwater images, the near-infrared images, and the fluorescence images. Specially, our algorithms provide an alternative way of thinking for researchers who don't want to follow the rules and ecosystems of machine learning, artificial intelligence, and neural networks. Most of the previously published works are related with machine learning, which needs a large database to be trained. However, our manuscript provides optional method for processing, which does not need a database. Moreover, our frameworks can raise the research interest about those scientists who are doing works about the image processing given that our designed filters can be environed to be integrated into several popular filters, including the Butterworth filter, the matched filter, and the filter based on watershed algorithm.

Research interest in seeking simple, affordable, and portable calculation platforms is growing, which is especially important in developing countries, resource-limited areas, and rural regions. It doesn’t need to use massive calculation systems for running our filters. No GPU or supercomputer or CPU array were used in our study. Only a Dell Desktop (OptiPlex 7070) is required for running our filters. This may raise the interest of the researchers in Asian, African, and Latin American countries who have limited access to computer resources.

Several simple functions including logarithm, hypotenuse, and hyperbolic secant are used. However, it does not mean that other functions are not workable. The selectivity of the functions need to be further studied in our future work.