Alves de Oliveira, Vinicius and Oberlin, Thomas and Chabert, Marie and Poulliat, Charly and Mickael, Bruno and Latry, Christophe and Carlavan, Mikael and Henrot, Simon and Falzon, Frederic and Camarero, Roberto Simplified entropy model for reducedcomplexity endtoend variational autoencoder with application to onboard satellite image compression. (2020) In: 7th International Workshop on OnBoard Payload Data Compression (OBPDC 2020), European Space Agency (ESA); Centre national d’études spatiales (CNES), 21 September 2020  23 September 2020 (Virtual, Greece).

(Document in English)
PDF (Author's version)  Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader 696kB 
Abstract
In recent years, neural networks have emerged as datadriven tools to solve problems which were previously addressed with modelbased methods. In particular, image processing has been largely impacted by convolutional neural networks (CNNs). Recently, CNNbased autoencoders have been successfully employed for lossy image compression [1,2,3,4]. These endtoend optimized architectures are able to dramatically outperform traditional compression schemes in terms of ratedistortion tradeoff. The autoencoder is composed of an encoder and a decoder both learned from the data. The encoder is applied to the input data to produce a latent representation with minimum entropy after quantization. The latent representation, derived through several convolutional layers composed of filters and activation functions, is multichannel (the output of a particular filter is called a channel or a feature) and nonlinear. The representation is then quantized to produce a discretevalued vector. A standard entropy coding method uses the entropy model inferred from the representation to losslessly compress this discretevalued vector. A key element of these frameworks is the entropy model. In earlier works [1,2,3], the learned representation was assumed independent and identically distributed within each channel and the channels were assumed independent of each other, resulting in a fullyfactorized entropy model. Moreover, a fixed entropy model was learned once, from the training set, preventing any adaptation to the input image during the operational phase. The variational autoencoder proposed in [4] proposed to use a hyperprior auxiliary network. This network estimates the hyperparameters of the representation distribution, for each input image. Thus, it does not require the assumption of a fullyfactorized model which conflicts with the need for context modeling. This variational autoencoder achieves compression performance close to the one of BPG (Better Portable Graphics) at the expense of a considerable increase in complexity.However, in the context of onboard compression, a tradeoff between compression performance and complexity has to be considered to take into account the strong computational constraints. For this reason, the CCSDS (Consultative Committee for Space Data Systems) lossy compression standard has been designed as a highly simplified version of JPEG2000. This work follows the same logic, however in the context of learned image compression. The aim of this paper is to design a simplified version of the variational autoencoder proposed in [4] in order to meet the onboard constraints in terms of complexity while preserving high performance in terms of ratedistortion. Apart from straightforward simplifications of the transform (e.g. reduction of the number of filters in the convolutional layers), we mainly propose a simplified entropy model that preserves the adaptability to the input image.A preliminary reduction of the number of filters reduces the complexity by 62% in terms of FLOPs with respect to [4]. It also reduces the number of learned parameters with a positive impact on the memory occupancy. The entropy model simplification exploits a statistical analysis of the learned representation for satellite images, also performed in [5] for natural images. This analysis reveals that most of the features are well fitted by centered Laplacian distributions. The complex hyperprior model based on a nonparametric distribution of [4] can thus be replaced by a simpler parametric centered Laplacian model. The problem then amounts to a classical and simple estimation of a single parameter referred to as the scale. Our simplified entropy models reduces the complexity of the variational autoencoder coding part by 22% and outperforms the endtoend model proposed in [1] for the high target rates.
Item Type:  Conference or Workshop Item (Paper) 

Audience (conference):  International conference proceedings 
Uncontrolled Keywords:  
Institution:  French research institutions > Centre National d'Études Spatiales  CNES (FRANCE) French research institutions > Centre National de la Recherche Scientifique  CNRS (FRANCE) Université de Toulouse > Institut National Polytechnique de Toulouse  Toulouse INP (FRANCE) Université de Toulouse > Institut Supérieur de l'Aéronautique et de l'Espace  ISAESUPAERO (FRANCE) Other partners > Thales (FRANCE) Université de Toulouse > Université Toulouse III  Paul Sabatier  UT3 (FRANCE) Université de Toulouse > Université Toulouse  Jean Jaurès  UT2J (FRANCE) Université de Toulouse > Université Toulouse 1 Capitole  UT1 (FRANCE) Other partners > ESA  ESTEC (NETHERLANDS) Other partners > Laboratoire de recherche en télécommunications spatiales et aéronautiques  TéSA (FRANCE) 
Laboratory name:  
Statistics:  download 
Deposited On:  31 Aug 2021 08:53 
Repository Staff Only: item control page