Research and Development of a Method for Transmitting Graphic Data over Low-speed Channels in Industrial IoT

In the field of deep learning, i.e. using convolutional neural networks (CNN), the developers managed to achieve significant success in image processing and machine vision. Despite this achievement, the approach is rarely applied when solving problems of image compression. The present study is devoted to this issue. A new compression model based on convolutional neural networks is proposed. Two CNNs are easily integrated into the end-to-end compression model. This allows achieving high-quality image compression with low data rates. The first type of CNN called “compact convolutional neural network” (ComCNN) studies the optimal compact representation of the input image. This representation stores structural information and is then encoded using codecs (for example, JPEG, JPEG2000, or BPG). The second type of CNN called “reconstruction convolutional neural network” (RecCNN) is used to restore the decoded image with high quality. To ensure the effective interaction of the two types of CNN, a single end-to-end learning algorithm is used for the simultaneous training of ComCNN and RecCNN models, which will simplify the accurate reconstruction of the decoded image using RecCNN. This development also allows to combine the proposed compression model with existing standards for image encoding. Keywords: deep learning, image compression, convolutional neural networks.