Improving Accuracy of Deep Learning-Based Compression Techniques by Introducing Perceptual Loss in Industrial IoT

Over the years there has been a great increase in the industrial Internet of things, and consequently, over 20 billion devices are now connected through the Internet. Internet of things is the connection of a wide variety of devices or entities over the Internet for storage and gathering of data. This will raise a need for compression techniques for remote sensing images, which is an enabling technology for the Internet of things. While various conventional technologies exist for the compression of remote sensing images, we require a low complexity technique that preserves the quality of the images without compromising over the various spatial parameters that these images possess. Deep learning methods through the use of convolution neural networks have proved to be a reasonable solution for the compression problem. The proposed method uses a deep learning-based method that uses autoencoders for encoding and decoding that incorporates perceptual loss to improve the accuracy of the reconstructed images. The loss functions we utilize are mean squared error loss, content loss, and style loss. We have also made use of lossless compression techniques using LZMA and deflate compression algorithms which do not result in any information loss. Experimental results show that our proposed compression method outperforms standard benchmark techniques.