Underwater images are often blurry or distorted because of the light attenuation or at times the back-scattering poorly affects the visibility. Researchers at Harbin Engineering University in China have discovered a solution to it by designing a machine learning algorithm that will create realistic underwater images. Another algorithm to give natural color to images and minimize the haze.
According to researchers, the quality and quantity of it will meet the modern standards and it is capable of processing upwards of 125 frames per second on a single graphic card.
Many of the image enhancement algorithms for underwater images are based on physical imaging models, which makes it unsuitable for the task. The new approach is related to GAN (generative adversarial network), an artificial intelligence (AI) model with a generator that tricks the discriminator to categorize synthetic samples as real-world samples. It produces a series of images of a particular survey site that are then sent to another algorithm, U-Net.
GAN was trained by the team on labeled scenes’ corpus having 3,733 images and equivalent depth maps, sea cucumbers, sea urchins and similar organisms that live indoor marine farms. Team also source opened data sets along with NYU Depth, which has thousands of underwater photographs.
After the training, researchers compared the photograph results of twin-model approach with baselines. The new approach delivered even color restoration. Without ruining the original input image’s underlying structure, the green-toned images were recovered accurately. The new approach also maintains the appropriate brightness and contrast of samples while recovering colors has unlike other solutions.
The method of reconstruction of frames from damaged footages is not used for the first time. Previously Cambridge Consultants’ DeepRay have also experimented with it to remove distortion due to opaque pane of glass. Also, Microsoft Research Asia also worked on an end-to-end system to independently perform video colorization.
Read next: The Future Of Artificial Intelligence In Retail
According to researchers, the quality and quantity of it will meet the modern standards and it is capable of processing upwards of 125 frames per second on a single graphic card.
Many of the image enhancement algorithms for underwater images are based on physical imaging models, which makes it unsuitable for the task. The new approach is related to GAN (generative adversarial network), an artificial intelligence (AI) model with a generator that tricks the discriminator to categorize synthetic samples as real-world samples. It produces a series of images of a particular survey site that are then sent to another algorithm, U-Net.
GAN was trained by the team on labeled scenes’ corpus having 3,733 images and equivalent depth maps, sea cucumbers, sea urchins and similar organisms that live indoor marine farms. Team also source opened data sets along with NYU Depth, which has thousands of underwater photographs.
After the training, researchers compared the photograph results of twin-model approach with baselines. The new approach delivered even color restoration. Without ruining the original input image’s underlying structure, the green-toned images were recovered accurately. The new approach also maintains the appropriate brightness and contrast of samples while recovering colors has unlike other solutions.
The method of reconstruction of frames from damaged footages is not used for the first time. Previously Cambridge Consultants’ DeepRay have also experimented with it to remove distortion due to opaque pane of glass. Also, Microsoft Research Asia also worked on an end-to-end system to independently perform video colorization.
Read next: The Future Of Artificial Intelligence In Retail
