Underwater image reconstruction using convolutional auto-encoder

One of the main tasks of AUVs is to capture deep-sea images like fishes, crabs, other living organisms and resources for information leading to research on deep-sea ecosystems. Acoustic transmission are used to establish wireless underwater communications between the AUV and the ship. However, there are some limitations in the communication channels due to limited bandwidth, multi-path, temperature distribution and change in the direction of transmitting source and receiving sensor which results in losses in data being transmitted. Initially, the captured images are enhanced to reduce the effect of light attenuation and then compressed for transmission through acoustic modems. Only an important part of image is being transmitted through set of data packets. The received data packets in the ship will be reconstructed to predict the presence of living organisms. The loss in data during transmission creates a difficulty for the operators to predict the exact information. In this research, to compensate this transmission loss, an efficient compression and reconstruction technique using convolutional autoencoder with minimal distortion is proposed. Finally, for evaluation of the proposed image compression technique, the quality of reconstruction of images with and without data loss will be compared using the quality metrics signal to noise ratio (PSNR), structural similarity index(SSIM) and perceptual quality of image.