ProgDTD: Progressive Learned Image Compression with Double-Tail-Drop Training

Ali Hojjat,Janek Haberer,Olaf Landsiedel

2024-04-30

Abstract:Progressive compression allows images to start loading as low-resolution versions, becoming clearer as more data is received. This increases user experience when, for example, network connections are slow. Today, most approaches for image compression, both classical and learned ones, are designed to be non-progressive. This paper introduces ProgDTD, a training method that transforms learned, non-progressive image compression approaches into progressive ones. The design of ProgDTD is based on the observation that the information stored within the bottleneck of a compression model commonly varies in importance. To create a progressive compression model, ProgDTD modifies the training steps to enforce the model to store the data in the bottleneck sorted by priority. We achieve progressive compression by transmitting the data in order of its sorted index. ProgDTD is designed for CNN-based learned image compression models, does not need additional parameters, and has a customizable range of progressiveness. For evaluation, we apply ProgDTDto the hyperprior model, one of the most common structures in learned image compression. Our experimental results show that ProgDTD performs comparably to its non-progressive counterparts and other state-of-the-art progressive models in terms of MS-SSIM and accuracy.

Computer Vision and Pattern Recognition,Image and Video Processing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to transform the existing non - progressive image compression methods into progressive image compression methods. Specifically, most of the existing image compression methods, whether traditional or learning - based, are non - progressive, which means that they require a complete compressed file to decode the image. This requirement is a challenge in many scenarios. For example, in the case of a slow network connection, it will cause the user or system to experience a delay before the image can be reconstructed for viewing or further processing. Progressive compression solves this problem, enabling the decoder to obtain an initial preview image even when receiving a small amount of data, and then reconstruct a higher - quality image by receiving more bits.

To achieve this goal, the paper introduces ProgDTD (Progressive Double - Tail - Drop Training), which is a method for converting non - progressive image compression methods into progressive ones. The core idea of ProgDTD is to modify the model during the training process so that it can sort the data in the bottleneck according to priority, thereby achieving progressive compression. This method does not require additional parameters and has a customizable progressive range.

The main contributions of the paper include:
1. Introducing ProgDTD, a method for making non - progressive models progressive.
2. ProgDTD is a training method that does not require additional parameters and is suitable for learning - based image compression.
3. ProgDTD has a customizable progressive range.
4. The performance of ProgDTD is comparable to that of the non - progressive version and other state - of - the - art benchmarks.

Through the experimental results, the paper shows that ProgDTD performs well in terms of MS - SSIM and accuracy. Although there is a slight decrease in PSNR, it can still generally provide performance comparable to that of non - progressive models. In addition, the paper also explores the performance of ProgDTD in different application scenarios, especially in the communication between bandwidth - limited edge devices and servers, where progressive image compression can significantly improve the efficiency of task completion.