Restauração de Imagens Subaquáticas

Diving into Clarity: Restoring Underwater Images using Deep Learning

Research work on underwater image restoration has been increasing in recent years and are extremely important for several applications in subaquatic robotics and scenarios. The acquisition process of good quality subaquatic images is a complex operation, representing a significant challenge for visual data capture and analysis, especially, due to the different underwater environments such as oceans, rivers and lakes. Distinct aspects contribute to the mentioned challenging acquisition, including i) water turbidity, caused by suspended particles; ii) presence of marine organisms, that contribute to the degradation of image quality and water scattering; and iii) uneven lighting and optical distortion, resulting in reduced visibility and loss of details.

A wide range of applications demand the understanding and processing of underwater images, especially using subaquatic robots. In an industrial context, different uses can be found in underwater robotics, offshore engineering and underwater exploration, which require restored images for object detection and identification, navigation and situational awareness. In a surveillance and monitoring context, the e hancement of underwater images, captured by subaquatic cameras and other sensors, enables the use of Remotely Operated Vehicle (ROV), assisting in wreck detection, threats and unauthorized objects and activities. Finally, in a marine ecology and biology context, it is paramount to comprehend and monitor underwater ecosystems, allowing the identification of species and their behavior.

Considering the underwater image acquisition complexity as mentioned earlier, the underwater image quality enhancement is a challenging problem, since there are many aspects affecting the subaquatic image quality. Additionally, the aforementioned real-world applications demonstrate the relevance of the addressed problem. 

In order to tackle this problem, advanced algorithms and techniques have been proposed to compensate the adverse effects of light scattering and absorption. By employing such techniques, underwater images can be transformed to reveal fine details, high contrast and precise color representation. Several methods transforming the image intensities can be employed such as color and gamma correction, histogram and contrast adjustment and unsharp enhancement. Figure 1 shows examples of raw underwater images, which were acquired in the ocean and in a river and their respective restored images.

In a previous work we have proposed an approach to underwater image quality enhancement, applying a fusion of color adjustment and compensation techniques, reaching a reasonable performance in terms of visibility quality and image quality metrics scores. However, in that work, we have used a simple strategy where the parameters were manually adjusted.

In this paper we propose a Deep Learning-based approach for underwater image restoration, enhancing the quality and clarity of subaquatic images, improving the navigation capability of subaquatic robots. We use a Convolutional Neural Network (CNN) regression model to estimate the best parameters to reduce the image degradation. Next a sequence of intensity transformation techniques are applied using the parameters found by the CNN, in order to improve the subaquatic images quality. Experiments were carried out using two well-known underwater image datasets. We have also carried out a comparison of our results with other relevant state-of-the-art algorithms, such as VRE, UDCP, IBLA, CBF, GDCP, UWCNN (UCNN), WaterNet (WN) and LI. The obtained results through qualitative and quantitative assessment metrics provide evidence supporting the effectiveness of the proposed approach for underwater image restoration, showcasing improved accuracy and significantly clearer underwater images. In the experiments, regarding the U45 and the UIEB datasets, seven state-of-the-art techniques were compared with our approach. For PSNR and SSIM quality metrics, in U45 dataset, the best results from comparison methods were 26.879 and 0.831, respectively. Meanwhile, our results were 26.967 and 0.847, respectively. For PSNR and SSIM quality metrics, in UIEB dataset, the best results from comparison methods were 27.157 and 0.788, respectively. Meanwhile, our results were 27.299 and 0.793, respectively. It is important to highlight that for PSNR and SSIM, the higher the value, the higher the restoration quality of underwater images.

Our work offers two main contributions, which can be summarized as follows:

• We propose a Deep Learning-based approach to learn the best parameters in the process of restoring the quality of underwater images. The proposed regression process acquires knowledge from different water conditions (like turbidity, low lighting, scattering and distortion in water), enabling the estimation of the better parameters for different underwater images. The proposed restoration approach presents high accuracy even regarding different water conditions. Thereby, our approach can improve the navigation capability of subaquatic robots in different underwater environments;

• We propose a challenging dataset composed by 276 underwater images, acquired in the Urubu river, a black-water tributary of the Amazon river. The proposed dataset comprises subaquatic images with intense turbidity and low lighting and scattering in the water. To the best of our knowledge this is the first dataset of underwater images from the Amazon region.

Publications:

MARTINHO, LAURA A. ; CALVALCANTI, JOÃO M. B. ; PIO, JOSÉ L. S. ; OLIVEIRA, FELIPE G. . Diving into Clarity: Restoring Underwater Images using Deep Learning. JOURNAL OF INTELLIGENT & ROBOTIC SYSTEMS, v. 110, p. 32, 2024.

MARTINHO, L. A. ; NETO, ODALISIO ; CALVALCANTI, JOÃO M. B. ; PIO, JOSÉ L. S. ; OLIVEIRA, FELIPE G. . An Approach for Fish Detection in Underwater Images. In: XVIII Workshop de Visão Computacional, 2023, São Bernardo do Campo/SP. Anais do XVIII Workshop de Visão Computacional. Porto Alegre, RS, Brasil: SBC, 2023. p. 6-11.

MARTINHO, LAURA A. ; OLIVEIRA, FELIPE G. ; CAVALCANTI, JOAO M. B. ; PIO, JOSE L. S. . Underwater image enhancement based on fusion of intensity transformation techniques. In: 2022 Latin American Robotics Symposium (LARS), 2022 Brazilian Symposium on Robotics (SBR), and 2022 Workshop on Robotics in Education (WRE), 2022, São Bernardo do Campo. 2022 Latin American Robotics Symposium (LARS), 2022 Brazilian Symposium on Robotics (SBR), and 2022 Workshop on Robotics in Education (WRE), 2022. v. 1. p. 348.

Acknowledgments:

This work was developed with support from Motorola, through the IMPACT-Lab R&D project at the Institute of Computing (IComp) of the Federal University of Amazonas (UFAM).

Team:

Laura A. Martinho, Graduate Student at Universidade Federal do Amazonas (UFAM)

João M. B. Calvalcanti, Associate Professor at Universidade Federal do Amazonas (UFAM)

José L. S. Pio, Associate Professor at Universidade Federal do Amazonas (UFAM)

Felipe G. Oliveira, Adjunct Professor at Universidade Federal do Amazonas (UFAM)