We propose a compression encoding method to perceptually optimize the image quality based on a novel quality metric, which emulates how the human visual system form opinion of a compressed image. Compared to the existing perceptual-optimized compression methods, which usually aim to minimize the detectability of compression artifacts and are sub-optimal in visually lossless regime, the proposed encoder aims to operate in the visually lossy regime. We implement the proposed encoder within the JPEG 2000 standard, and demonstrate its advantage over both detectability-based and conventional MSE encoders.

Traditional image compression methods primarily focus on maximizing the fidelity of the compressed image using image quality driven distortion metrics, which are ideally suited for human observers but are not necessarily optimal for machine observers, i.e., automated image exploitation algorithms. For machine observers, task-based distortion metrics, such as probability of error, have been shown to be more effective for tasks such as object detection and classification. This motivates an approach to a task-based image compression, within the JPEG 2000 framework, which preserves the information that is most relevant for the given task. Our proposed method produces a JPEG 2000 compliant compressed codestream, which can be decoded by any JPEG 2000 compliant decoder. We demonstrate the feasibility and the effectiveness of our task-based image compression approach on a simple object classification and detection problem and quantify its performance relative to a conventional MSE encoder.

In this paper, we propose a visual discrimation model to enable perceptually optimized JPEG200 compression for both near-threshold and supra-threshold regimes. The performance of the proposed approach is validated by comparing it to the conventional Mean Squared Error (MSE)-optimized compression.

Image compression systems that exploit the properties of the Human Visual System (HVS) have been studied extensively over the few decades. For the JPEG2000 image compressionstandard, several methods to optimize perceptual quality have been proposed. In 2013, Han et al. proposed a visually lossless compression approach based on the irreversible pipeline defined in the JPEG2000 standard. In this approach, visibility thresholds were measured using psychovisual experiments. These thresholds were then incorporated in a JPEG2000 encoder to ensure that quantization distortions remain below visible levels in the compressed codestreams. In this work, we investigate the use of a similar approach for the reversible pipeline of the JPEG2000 standard. Our motivation is to allow the creation of a scalable codestream that can provide both visually lossless and numerically lossless representations from a single codestream. By comparing the difference in compression performance between the reversible and irreversible pipelines, we also quantify the overhead associated with the reversible pipeline for visually lossless compression.

Estimating the angular separation between two incoherently radiating monochromatic point sources is a canonical toy problem to quantify spatial resolution in imaging. In recent work, Tsang et al. showed, using a Fisher Information analysis, that Rayleigh's resolution limit is just an artifact of the conventional wisdom of intensity measurement in the image plane. They showed that the optimal sensitivity of estimating the angle is only a function of the total photons collected during the camera's integration time but entirely independent of the angular separation itself no matter how small it is, and found the information-optimal mode basis, intensity detection in which achieves the aforesaid performance. We extend the above analysis, which was done for a Gaussian point spread function (PSF) to a hard-aperture pupil proving the information optimality of image-plane sinc-Bessel modes, and generalize the result further to an arbitrary PSF. We obtain new counterintuitive insights on energy vs. information content in spatial modes, and extend the Fisher Information analysis to exact calculations of minimum mean squared error, both for Gaussian and hard aperture pupils.

The goal of this paper is to present a non-iterative and more importantly an extremely fast algorithm to reconstruct images from compressively sensed (CS) random measurements. To this end, we propose a novel convolutional neural network (CNN) architecture which takes in CS measurements of an image as input and outputs an intermediate reconstruction. We call this network, ReconNet. The intermediate reconstruction is fed into an off-the-shelf denoiser to obtain the final reconstructed image. On a standard dataset of images we show significant improvements in reconstruction results (both in terms of PSNR and time complexity) over state-of-the-art iterative CS reconstruction algorithms at various measurement rates. Further, through qualitative experiments on real data collected using our block single pixel camera (SPC), we show that our network is highly robust to sensor noise and can recover visually better quality images than competitive algorithms at extremely low sensing rates of 0.1 and 0.04. To demonstrate that our algorithm can recover semantically informative images even at a low measurement rate of 0.01, we present a very robust proof of concept real-time visual tracking application.

Compression is a key component in many imaging systems in order to accommodate limited resources such as power and bandwidth. Image compression is often done independent of the specific tasks that the systems are designed for, such as target detection, classification, diagnosis, etc. Standard compression techniques are designed based on quality metrics such as mean-squared error (MSE) or peak signal to noise ratio (PSNR). Recently, a metric based on task-specific information (TSI) was proposed and successfully incorporated into JPEG2000 encoding. It has been shown that the proposed TSI metric can optimize the task performance. In this work, a joint metric is proposed to provide a seamless transition between the conventional quality metric MSE and the recently proposed TSI. We demonstrate the effectiveness and flexibility of the proposed joint TSI metric for target detection tasks. Furthermore, it is extended to video tracking applications to demonstrate the robustness of the proposed metric. Experimental results show that although the metric is not directly designed for the applied task, better tracking performance can still be achieved when the joint metric is used, compared to results obtained with the traditional MSE metric.

We present a compressive imager demonstrator based on a scalable, parallel architecture. It primarily utilizes information-optimal projections and a Piece-wise Linear Minimum Mean Square Error Estimator (PLE-MMSE) combined with a block-based statistical model of natural images. Such system delivers high-resolution images from low resolution sensor with near real-time snapshots. This testbed provides a highly programmable compressive imager that allows testing of a variety of projection designs for different tasks (e.g. random binary, PCA) and also enables adaptive or dynamic designs.

Compression is a key component in imaging systems hat have limited power, bandwidth, or other resources. In many applications, images are acquired to support a specific task such as target detection or classification. However, standard image compression techniques such as JPEG or JPEG2000 (J2K) are often designed to maximize image quality as measured by conventional quality metrics such as mean-squared error (MSE) or Peak Signal to Noise Ratio (PSNR). This mismatch between image quality metrics and ask performance motivates our investigation of image compression using a task-specific metric designed for the designated tasks. Given the selected target detection task, we first propose a metric based on conditional class entropy. The proposed metric is then incorporated into a J2K encoder to create compressed codestreams that are fully compliant with the J2K standard. Experimental results illustrate that the decompressed images obtained using the proposed encoder greatly improve performance in detection/classification tasks over images encoded using a conventional J2K encoder.