Alternative Underwater Image Restoration Based on Unsupervised Learning and Autoencoder with Degradation Block

• Claudio Mello Jr. FURG
• Bryan Moreira FURG
• Paulo Drews-Jr FURG
• Silvia Botelho FURG

Resumo

Underwater image restoration represents a challenge to computer vision and image processing based on machine learning. Recent methodologies to tackle this problem are based on learning. However, the lack of paired image datasets leads the researches to synthesize datasets. We present a new unsupervised learning algorithm that no requires paired dataset to train an encoder-decoder-like neural network for underwater image restoration. Encoder-decoder network learns to represent its input data in a latent representation and reconstruct then in the output. During the training stage, our algorithm applies the output image to a degradation block based on image formation model that reinforces its degradation. The degraded and input images are matched using a loss function. After the training process, we are able to obtain a restored image from the decoder. We focus on underwater inspection and our method relies on small dataset and a light neural network. Underwater images are used to evaluate and validate our algorithm. The qualitative and quantitative results show the improvement provided by our method.

Palavras-chave: Image restoration, Training, Degradation, Neural networks, Image color analysis, Proposals, Unsupervised learning