In this project, you will design a post-filtering scheme for the recently proposed direction-adaptive partitioned block transform (DA-PBT) for images [2]. The DA-PBT exploits directionality in images to improve the coding performance with non-rectangular partitions and directional transforms. Therefore, unexplored coding artifacts due to the novel partitioning scheme may be introduced. You will investigate potential post-filtering schemes to reduce the artifacts with or without side information, such as the block modes and directions chosen in the DA-PBT, and statistical information between the original and the reconstructed images.

For compression applications, the coefficients obtained through SVD or QR factorization have to be efficiently entropy-coded. The goal of this project is to develop entropy coding schemes for efficient
critical representation of the detail signal of the LP. Alternatively, other transforms for critical representation of the detail signal and their corresponding entropy coding schemes can be explored.

In this project, you will develop algorithms for the estimation of the PSNR of a JPEG-compressed image without reference to the original, under various conditions. In the basic scenario, you will assume that the estimator has access to the JPEG bitstream, which includes the quantization matrix. The problem is more challenging if the estimator only knows the reconstructed image, the quantization information having been lost. A further twist would allow for the original image to be compressed, reconstructed and compressed again at a different quality. Further variations are possible: how can you improve the estimator if it additionally has access to a thumbnail version (uncompressed or compressed) of the original image?

One paper (of many) that discusses statistical analysis of discrete cosine transform (DCT) coefficients for PSNR estimation is [1]. The methods in [2] may be useful for recovering lost quantization information.

For this project, students should investigate lossy feature compression techniques. Lossy compression techniques should exploit correlations between the different dimensions of feature vectors. A rate-distortion-optimized Lagrangian framework should be developed for compressing the features. Initially, mean-squared error (MSE) can be used as the distortion metric. Other distortion metrics like the L1 norm and Mahanalobis distance should also be considered. Additionally, the impact of compression on the classification rate should be studied.

Feature data sets and feature matching software will be provided at the onset of the project.

For this project, students should extract features from (i) uncompressed images and (ii) JPEG-compressed [3] images, and explore how feature quality is affected by JPEG compression. Feature quality is evaluated in terms of the image matching accuracy achievable with the (possibly degraded) features. How does changing the JPEG coding parameters, such as the quantization matrix, affect feature quality? How do JPEG compression artifacts affect feature quality? It is desirable to optimize the JPEG coding parameters to aid feature extraction, as opposed to the original goal of optimizing coding parameters for perceptual image viewing. Since images stored on digital cameras are JPEG-compressed, this project can lead to a new set of JPEG settings that are appropriate for feature extraction rather than typical image viewing.

Feature data sets (e.g. for buildings or for CD covers), feature extraction software, and feature matching software will be provided at the onset of the project.

In this project, you will develop algorithms that only observe the edited image in order to distinguish between different types of image forgery and to determine the parameters of the forgery method (such as the scaling factor of a resize operation). The task is more challenging if the image is observed after another round of JPEG compression, as would be typical.

A general framework for forgery detection with intrinsic fingerprints is described in [1]. The references in [1] may also useful; for example, [2] discusses a histogram-matching method to discover the original quantization matrix in twice JPEG-compressed images.