Researchers at Harbin Engineering University design an AI algorithm for underwater image dehazing and colorization.

2020-01-01

We now see that the underwater images we see are blurry and distorted, because phenomena such as light attenuation and backscatter have a detrimental effect on visibility. To solve this problem, many researchers and scholars have made efforts. Cambridge Consultants' DeepRay uses a GAN trained on a dataset of 100,000 static images to eliminate distortion caused by opaque glass plates. The open-source DeOldify project uses a series of AI models, including GAN, to colorize and restore old images and films. In September, scientists at Microsoft Research Asia detailed an end-to-end system for automatic video coloring. Last year, Nvidia researchers described a framework that can infer colors from a colored and annotated video frame. In June, Google AI launched an algorithm that can colorize grayscale videos without human supervision.


However, researchers at Harbin Engineering University recently contributed different results. They designed two machine learning algorithms, one that can generate realistic underwater images and a second algorithm that can train on these images to restore natural colors and reduce haze. They say their method matches the latest technology in terms of quality and quantity, and can process up to 125 frames per second on a single graphics card.


The team pointed out that most underwater image enhancement algorithms (such as those that adjust white balance) are not based on physical imaging models, which makes them unsuitable for other tasks. In contrast, this approach utilizes a Generative Adversarial Network (GAN) (an AI model consisting of a generator that tries to deceive a discriminator into classifying synthetic samples as real samples) to generate a set of images of specific survey locations, and introduces a second algorithm called U-Net.The team trained the GAN on a labeled scene corpus, which included 3,733 images and corresponding depth maps, mainly of scallops, sea cucumbers, sea urchins, and other such organisms living in indoor marine farms.


After training, the researchers compared the results of their dual-model approach with the baseline results. They noted that this technology has the following advantages: uniform color restoration and the ability to restore images with a green hue without compromising the structure of the original input image. Typically, it can also restore color while maintaining "appropriate" brightness and contrast, which other solutions are not particularly adept at.

RECENT AI NEWS

RECENT AI TOOLS