Underwater image enhancement using non-dynamic stochastic resonance

The aim of this work is to study and apply the phenomenon of stochastic resonance (SR), which occurs in nonlinear systems across various disciplines such as physics, chemistry, biology, and others. This research explores a method for enhancing the quality of low-contrast images based on this phenomenon. By the beginning of this study, there were already many successful works on this topic. However, they were implemented mostly from a technical point of view, relying on dry mathematical methods. This thesis offers, so to speak, a "physicist's view". Parallels are drawn between the processes occurring in computational systems during image processing and physical systems, surprising analogies are discovered, and new solutions are proposed based on them. Initially, we chose this method for processing and preparing a dataset of images for subsequent deep learning analysis. However, upon studying and implementing it, we encountered the challenge that obtaining good results requires careful selection of the optimal operating parameters, which consumes the researcher's time and resources. During our research, we found solutions to improve and automate the method, which we would like to outline in this work. As it is known, non-dynamic SR is a threshold phenomenon. The complexity lied in determining the threshold value and finding the optimal standard deviation for the added noise, as they are functions of multiple variables and are derived through complex mathematical formulas. However, in our work, we simplified this process. All the necessary parameters were extracted from the input data. We took the mean value of the input pixels as the threshold value, with the condition that the mean value of the added noise is zero. Furthermore, according to the concept of resonance, for resonance to occur, the system's frequency must match with the frequency of the external force. We realized that in the case of images, the role of frequency is played by the standard deviation. Therefore, in order to achieve resonance (in our case, the transition of the system to either low-contrast or high-contrast states), we generated noise with a standard deviation equal to the standard deviation of the input image. These improvements allow us to achieve the best result without any manual intervention and in just one iteration. By calculating the necessary parameters from the image itself, our approach tailors the enhancement process to the unique characteristics of each image, thereby circumventing the limitations of traditional denoising and enhancing methods. Additionally, the proposed ideas offer the ability to focus on any region of the image and detect objects. In other words, the possibility of SR-zoom is implemented, which is shown in the examples. The study presents practical solutions that allow researchers to save time and resources in image processing. Before moving on to specific images, we demonstrated the effectiveness of our approach using the well-known Lena image. Further, all the solutions have been implemented using underwater images, including UIEB dataset, which is a sought-after and attractive research direction. Underwater image processing represents a fundamental component of marine engineering, essential for enhancing the quality of underwater images afflicted by blurriness, incompleteness, or various artifacts. The challenges posed by such imperfections necessitate the development of robust image enhancement techniques. While a plethora of methods for image enhancement exist in the literature, ranging from traditional approaches to the techniques on the ML or DL methods, a notable limitation arises with many ML and DL techniques. These approaches often exhibit a propensity to memorize data directly from the training dataset, rather than generalize from it. Consequently, while these methods may demonstrate optimal performance in controlled laboratory environments, they frequently encounter significant accuracy loss when deployed in real-world scenarios. Denoising emerges as a pivotal aspect of traditional image enhancement algorithms, typically accomplished through the introduction of random threshold values and noise to the input image. However, this conventional denoising approach may inadvertently lead to degradation in accuracy and loss of image quality. In response to these challenges, this study proposes a novel approach to denoising, leveraging the principles of SR. Evaluations for underwater image enhancement can be divided into two main categories: automatic evaluation metrics and the human visual system. The automatic evaluations utilize six metrics. Among these, four are widely used in image enhancement and restoration problems, namely PSNR (peak signal-to-noise ratio), MSE (mean squared error), SSIM (structural similarity index measure), and PCQI (patch-based contrast quality index). The other two metrics are specifically designed for underwater image enhancement: UCIQE (underwater colour image quality evaluation) and UIQM (underwater image quality measure). To ensure underwater image enhancement accuracy, these kinds of different evaluation metrics should be discussed including their limitations and reliability, and have also provided a report on human visual evaluation, highlighting its importance. We provide a comparative discussion of various image enhancement methods, evaluated using the full-reference image quality metrics on the UIEB dataset. The methods under comparison include Fusion-based, Retinex-based, GDCP, Histogram prior, Blurriness-based, Water CycleGAN, Dense GAN, Water-Net, and our proposed method. The metrics used for evaluation are MSE, accuracy, PSNR and SSIM. Our proposed method outperforms all other methods across all evaluation metrics. Specifically, it achieves the lowest MSE of 0.97%, which is significantly lower than the next best method, Water-Net, which has an MSE of 1.23%. This indicates a higher precision in image reconstruction with our method. The accuracy of our method is also the highest at 99.03%, compared to Water-Net's 98.77% and the Fusion-based method's 98.27%. This high accuracy demonstrates the robustness of our method in preserving the correct details of the underwater images. In terms of PSNR, our method achieves the highest value of 20.23 dB. This is a substantial improvement over Water-Net's 19.11 dB and the Fusion-based method's 17.61 dB, highlighting our method's superior noise reduction capabilities. The SSIM index, which evaluates the visual impact of three characteristics of an image: luminance, contrast, and structure, is highest for our method at 0.8000. This is slightly better than Water-Net's 0.7971, but significantly higher than the SSIM values of other methods such as Dense GAN (0.4426) and GDCP (0.5121). The high SSIM score reflects the effectiveness of our method in maintaining the structural integrity and overall visual quality of the images. Among the traditional and machine learning-based methods, the Fusion-based method performs relatively well, but it still falls short compared to our method. The Retinex-based method, although commonly used for image enhancement, shows lower performance metrics. Methods like GDCP and Water CycleGAN demonstrate poorer performance, particularly in terms of MSE and PSNR, indicating less effective noise reduction and detail preservation. The Water-Net method shows strong results across most metrics but is still surpassed by our method in every category. Our proposed method achieves the highest scores also in no-reference UCIQE and UIQM metrics, with 0.6791 and 1.34, respectively. By significantly enhancing both of these metrics scores, our method sets a new benchmark for underwater image enhancement, surpassing both traditional and contemporary techniques. The improvements can be attributed to the effective application of SR, which leverages the unique characteristics of each image to optimize the enhancement process. This approach not only addresses the noise and blurriness typically associated with underwater images but also improves colour balance and overall perceptual quality. The comparative analysis clearly demonstrates that our proposed method achieves the best overall performance across all key image quality metrics, making it a highly effective solution for underwater image enhancement. These findings highlight that our SR-based approach is highly effective for underwater vision, successfully removing "water noise" and delivering superior accuracy and structural preservation. In the future, this technique can form the basis for an advanced underwater vision system with the ability to zoom in and detect various objects, enhancing its practical applications in marine engineering and underwater exploration. While our proposed method demonstrates superior performance in underwater image enhancement, it is not without limitations. Although our method effectively tailors the enhancement process to individual images, it may not generalize well to drastically different underwater conditions or image qualities not represented in the UIEB dataset. Additionally, the computational complexity associated with calculating image-specific parameters using stochastic resonance can be resource-intensive, potentially limiting real-time applications. Future work should focus on optimizing the computational efficiency of the method to facilitate real-time processing. Moreover, expanding the dataset to include a broader range of underwater environments nd conditions would enhance the generalizability of the method. Another avenue for future research could involve integrating advanced machine learning techniques to further refine the enhancement process, possibly combining the strengths of traditional and modern approaches.