Deep Learning-Based Adaptive Image Compression System for a Real-World Scenario

Deep Learning-based (DL) image compression has shown prominent results compared to standard image compression techniques like JPEG, JPEG2000, BPG and WebP. Nevertheless, neither DL nor standard techniques generally can cope with critical real-world scenarios, with stringent performance constraints. In order to explore the nature of this gap, we first introduce an industrial scenario, which contemplates real-time compression of high-resolution images, with strict requirements on a number of quality-performance indicators, namely: the output image quality, the hardware, and the compression complexity. Next, we propose a DL-based image compression model, i.e. a Convolutional Residual Autoencoder (CRAE). In particular, CRAE integrates some structural benefits of a deep neural network, including PReLU activation function and sub-pixel convolution, which have proven to be especially suitable for image compression tasks. We analyze the performance of the proposed CRAE approach by adopting two types of processing: (i) global and, (ii) patch-based processing of image data. To test the models, we exploit a dataset composed of high-resolution images provided by the MERMEC company composed of consecutive images of the railway track captured by a machine vision system called V-CUBE. Furthermore, the company provided strict compression requirements that needed to be met by the developed system. Preliminary results of an ongoing study indicates that the proposed image compression system can meet the requirements by MERMEC with reasonable performance, with a mild advantage observed for full-based CRAE. The obtained outcomes suggests that CRAE can adapt to the specific structure of the given dataset and extracts the salient recurrent patterns inside an image. In summary, this line of research represents the core of the future plug-and-play DL architecture for constrained image compression.