Quantization (image processing)

About: Quantization (image processing) is a research topic. Over the lifetime, 7977 publications have been published within this topic receiving 126632 citations.

Quantization table adjustment

Abstract: The method for adjusting quality during image capture includes computing a discrete cosine transform of a digital image to create DCT coefficients. A quantization table is generated that specifies frequency bands to be filtered and the DCT coefficients are digitized using the quantization table. It is preferred that the DCT coefficients be ordered in a zig-zag sequence to facilitate run-length encoding.

Image Processing Apparatus, Image Processing Method, and Computer Program

Abstract: An image processing apparatus for generating a motion compensation image. The image processing apparatus includes: a pixel-relative-position calculation section calculating a relative position between a pixel position of a pixel constituting the motion compensation image and a pixel position of a reference pixel of a reference image to be used for calculating a pixel value of the pixel on the basis of motion vector information; a relative-position quantization section performing quantization processing of relative-position information calculated by the pixel-relative-position calculation section and generating quantized-relative-position information; and a motion-compensation-image generation section generating a motion compensation image by calculating a pixel value of a constituent pixel of the motion compensation image on the basis of the quantized-relative-position information and a pixel value of the reference pixel.

On face detection in the compressed domain

Abstract: We propose a fast face detection algorithm that works directly on the compressed DCT domain. Unlike the previous DCT domain processing designs that are mainly based on skin-color detection, our algorithm analyzes both color and texture information contained in the DCT parameters, therefore could generate more reliable detection results. Our texture analysis is mainly based on statistical model training and detection. A number of fundamental problems, e.g., block quantization, preprocessing in the DCT domain, and feature vector selection and classification in the DCT domain, are discussed.

A new approach for JPEG resize and image splicing detection

Abstract: Today's ubiquitous digital media are easily tampered by, e.g., removing or adding objects from or into images without leaving any obvious clues. JPEG is a most widely used standard in digital images and it can be easily doctored. It is therefore necessary to have reliable methods to detect forgery in JPEG images for applications in law enforcement, forensics, etc. In this paper, based on the correlation of neighboring Discrete Cosine Transform (DCT) coefficients, we propose a method to detect resized JPEG images and spliced images, which are widely used in image forgery. In detail, the neighboring joint density features of the DCT coefficients are extracted; then Support Vector Machines (SVM) are applied to the features for detection. To improve the evaluation of JPEG resized detection, we utilize the shape parameter of generalized Gaussian distribution (GGD) of DCT coefficients to measure the image complexity.The study shows that our method is highly effective in detecting JPEG images resizing and splicing forgery. In the detection of resized JPEG image, the performance is related to both image complexity and resize scale factor. At the same scale factor, the detection performance in high image complexity is, as can be expected, lower than that in low image complexity.

Efficient Super Resolution Using Binarized Neural Network

Abstract: Deep convolutional neural networks (DCNNs) have recently demonstrated high-quality results in single-image super-resolution (SR). DCNNs often suffer from over-parametrization and large amounts of redundancy, which results in inefficient inference and high memory usage, preventing massive applications on mobile devices. As a way to significantly reduce model size and computation time, binarized neural network has only been shown to excel on semantic-level tasks such as image classification and recognition. However, little effort of network quantization has been spent on image enhancement tasks like SR, as network quantization is usually assumed to sacrifice pixel-level accuracy. In this work, we explore an network-binarization approach for SR tasks without sacrificing much reconstruction accuracy. To achieve this, we binarize the convolutional filters in only residual blocks, and adopt a learnable weight for each binary filter. We evaluate this idea on several state-of-the-art DCNN-based architectures, and show that binarized SR networks achieve comparable qualitative and quantitative results as their real-weight counterparts. Moreover, the proposed binarized strategy could help reduce model size by 80% when applying on SRResNet, and could potentially speed up inference by 5 times.