SEMANTICALLY AWARE NEURAL NETWORK IMAGE COMPRESSION FOR UNDERWATER SURVEILLANCE

Acoustic transmission is used to establish wireless underwater communication between an underwater vehicle and the control station, where the operator is located. However, there are some limitations in the communication channels, due to limited bandwidth, which results in transmission data losses. For real-time applications, these losses interfere with the understanding of the events. In underwater missions, images are usually stored with little possibility of use due to their non-variational characteristics. Determining the appropriate level of compression applied, according to the amount of semantic content of a certain image, is an alternative approach. Therefore, in this work, encoding underwater images using a Variational Autoencoder is proposed. The main advantages of this image compression technique are its low distortion and high bit rate. The authors of this study also discuss the use of SIFT as a metric attempt to quantify the variation of information in an image.