Neural Network Based Image Compression Research Articles

Toward Robust Neural Image Compression: Adversarial Attack and Model Finetuning

Deep neural network-based image compression has been extensively studied. However, the model robustness which is crucial to practical application is largely overlooked. We propose to examine the robustness of prevailing learned image compression models by injecting negligible adversarial perturbation into the original source image. Severe distortion in decoded reconstruction reveals the general vulnerability in existing methods regardless of their settings (e.g., network architecture, loss function, quality scale). A variety of defense strategies including geometric self-ensemble based pre-processing, and adversarial training, are investigated against the adversarial attack to improve the model's robustness. Later the defense efficiency is further exemplified in real-life image recompression case studies. Overall, our methodology is simple, effective, and generalizable, making it attractive for developing robust learned image compression solutions. All materials are made publicly accessible at https://njuvision.github.io/RobustNIC for reproducible research.

Recurrent neural network based image compression

Recurrent neural network based image compression

Deep Image Compression in the Wavelet Transform Domain Based on High Frequency Sub-Band Prediction

In this paper, we propose to use deep neural networks for image compression in the wavelet transform domain. When the input image is transformed from the spatial pixel domain to the wavelet transform domain, one low-frequency sub-band (LF sub-band) and three high-frequency sub-bands (HF sub-bands) are generated. Low-frequency sub-band is firstly used to predict each high-frequency sub-band to eliminate redundancy between the sub-bands, after which the sub-bands are fed into different auto-encoders to do the encoding. In order to further improve the compression efficiency, we use a conditional probability model to estimate the context-dependent prior probability of the encoded codes, which can be used for entropy coding. The entire training process is unsupervised, and the auto-encoders and the conditional probability model are trained jointly. The experimental results show that the proposed approach outperforms JPEG, JPEG2000, BPG, and some mainstream neural network-based image compression. Furthermore, it produces better visual quality with clearer details and textures because more high-frequency coefficients can be reserved, thanks to the high-frequency prediction.

Neural Network based Image Compression for Memory Consumption in Cloud Environment

Background/Objectives: The Main aim of this Hybrid Image compression method is that it should provide good picture quality, better compression ratio and also it can remove block artifacts in the reconstructed image. Methods/Statistical analysis: To compress an image using the proposed algorithm, images are first digitized. With the digital Information of an image different types of transformations are applied. In this method wavelet transformations (haar, daubechies wavelet transformations) are used. The output of transformation coefficients are quantized into nearest integer values. Vector Quantization takes an important role in quantizing the transformation coefficients. After quantization they are encoded by using any one of the compression encoding techniques. Huffmann encoding is used for compressing Tablet images and Tablet strip images. It is derived from exact frequencies of text. The variable length code table is an output of Huffmann’s algorithm. The source symbol is encoded and stored in the above table which is further transferred through the channel for decoding. Findings: Since Unsupervised Neural Network learning algorithms are added in this algorithm increase the picture quality is improved and it removes the problems of block artifacts. Conclusion/Application: Since cloud computing provides elastic services, high performance and scalable large data storage, to facilitate long term storage and efficient transmission Image files are compressed and stored using this hybrid compression algorithm to enhance the performance of recent compression algorithms. The compressed and reconstructed images are evaluated using the error measures like CR (Compression Ratio), PSNR (Peak Signal Ratio). It shows that the above explained algorithm provides better results than the traditional results.

Neural Network Based Image Compression Approach to Improve the Quality of Biomedical Image for Telemedicine

Neural Network Based Image Compression Approach to Improve the Quality of Biomedical Image for Telemedicine

Embedded Image Compression Algorithm and FPGA Implementation Based on BP Neural Networks

At present, because more embedded image compressions are single, various compression methods have not transplant to embedded equipment. In this paper, A BP neural network based image compression methods have been proposed. The neural network is trained more and more, and obtained a set of weights and thresholds usefully. Then, use the FPGA to achieve, In the FPGA, using the framework of soft-core Nios Ⅱ way. Ultimately, compression program written using Verilog and burned into the FPGA. Experiments show that the system has the advantages of high compression ratio, small size, and can stable operation.

ASIC Implementation of Neural Network Based Image Compression

ASIC Implementation of Neural Network Based Image Compression

Compression of digital chest radiographs with a mixture of principal components neural network: evaluation of performance.

The performance of a new, neural network-based image compression method was evaluated on digital radiographs for use in an educational environment. The network uses a mixture of principal components (MPC) representation to effect optimally adaptive transform coding of an image and has significant computational advantages over other techniques. Nine representative digital chest radiographs were compressed 10:1, 20:1, 30:1, and 40:1 with the MPC method. The five versions of each image, including the original, were shown simultaneously, in random order, to each of seven radiologists, who rated each one on a five-point scale for image quality and visibility of pathologic conditions. One radiologist also ranked four versions of each of the nine images in terms of the severity of distortion: The four versions represented 30:1 and 40:1 compression with the MPC method and with the classic Karhunen-Loève transform (KLT). Only for the images compressed 40:1 with the MPC method were there any unacceptable ratings. Nevertheless, the images compressed 40:1 received a top score in 26%-33% of the evaluations. Images compressed with the MPC method were rated better than or as good as images compressed with the KLT technique 17 of 18 times. Four of nine times, images compressed 40:1 with the MPC method were rated as good as or better than images compressed 30:1 with the KLT technique.

A neural network based image compression system

It is proposed that vector quantization be implemented for image compression based on neural networks. Separate codebooks for edge and background blocks are designed using Kohonen (1984) self-organizing feature maps to preserve edge integrity and improve the efficiency of codebook design. A system architecture is proposed, and satisfactory performance is achieved. >