Region-of-Interest Data Compression with Prioritized Buffer Management

We describe an integrated system for intelligent compression and transmission of copious data acquired by spaceborne instruments. At its core, our system contains a modification of a progressive image compression algorithm, ICER, that will be used on the Mars Exploration Rovers (to be launched in 2003). The ICER algorithm applies a wavelet decomposition and prioritizes the compressed bit layers from the wavelet subbands so as to (attempt to) progressively transmit the layer that gives the largest improvement in image quality per transmitted bit. Our modified version accepts additional input priorities that reflect the relative importance of various “regions of interest” in the source data, and arranges its output packets to reflect both the input regional priorities and the wavelet bit layer priorities. The output of the data compression module is supervised by an intelligent buffer manager that shuffles the prioritized packets from many different source images and tries to select packets for transmission that will maximize the total science value received on the ground. Just as importantly, it attempts to discard only the least valuable packets when the buffer overflows, which is inevitable if the average data transmission rate is lower than the average data collection rate. After briefly describing our current system, the paper analyzes in more detail various algorithms for optimization of the buffer control algorithm, for bit allocation across regions of the source data, and for efficiently compressing data features useful for establishing a prioritization feedback loop between spacecraft and ground. I. DESCRIPTION OF THEOVERALL SYSTEM We are developing integrated data compression and buffer management algorithms to maximize the science value of data returned from spacecraft instruments. Typically, imagers and remote sensors have the capability to collect far more data than can be transmitted to earth, and it is important to maximize the science value of the data returned. Onboard science processing algorithms that recognize scientifically relevant features in the collected data can be used to drive progressive data compression algorithms such as wavelet-based image compression. During progressive compression, the science data is parsed into hierarchical data segments that yield continual but diminishing improvement of fidelity with each segment. The compression schemes for science-directed progressive compression produce This work was funded by the ESTO Technology Program and performed at the Jet Propulsion Laboratory, California Institute of Technology, and at the Signal and Image Processing Institute, Integrated Media Systems Center, University of Southern California. data segments specially tailored to “regions of interest” (ROIs) specified by science processing modules. The prioritized buffer manager tries to ensure that the highest priority data segments are transmitted first, and the least valuable data segments are discarded. Our approach is to adapt existing progressive compression algorithms for amenability with identified ROIs, and to develop buffer strategies for prioritizing, storing, and delivering the most valuable compressed segments, and eventually reconstituting the original data. We attempt to incorporate ROI considerations across many images or different data types. These algorithms are subject to practical limits on the onboard computers speed, memory, and storage. We attempt to measure the gain in science return versus required processing speed, memory, and storage of the onboard computer. II. T HE CURRENT ALGORITHMS The current system takes the raw data together with a data prioritization map and produces packets of compressed data using a modified version of the ICER wavelet-based compression software. 1 Then the prioritized buffer manager decides which packets to admit, discard, or transmit, depending on their priorities and given constraints on the buffer’s input and output rates. A. DATA PRIORITIZATION A data prioritization map is an assignment of a priority number to each pixel of an image. In our implementation, a priority number is an integer in the range 0–5, with higher numbers indicating higher priority. A difference of some number b between two priority numbers indicates that the higher priority pixel should be reconstructed to roughlyb more bits of precision than the lower priority pixel. The priority map is generated by identifying and classifying features of the source image that are of interest to the data’s scientistusers. How to best produce such a map is a very complex question, and one best left to the scientists for each intended application. Our system does include some rudimentary general-purpose classification algorithms based on colors and textures, but the main purpose of this system is to provide scientists with powerful tools for intelligently compressing and transmitting their source data no matter how the priorities have been identified. B. ROI COMPRESSION ALGORITHM We have modified the ICER software to accept the additional input of a priority map indicating the relative importance of regions of 1Developed at JPL by A. Kiely and M. Klimesh. the image. The compression takes the priority map into account and produces as output a sequence of packets paired with priority values. The priority values serve as a means of comparing the importance of packets, even those from separate images (assuming of course that the corresponding priority maps are also comparable). As in the basic ICER, the input image is transformed using a wavelet decomposition. A priority map is computed for the wavelettransformed image such that the spatial correspondence that existed between the original image and priority map is maintained between the transformed image and transformed priority map. Specifically, the priority corresponding to a transformed pixel is equal to the maximum priority of the original image pixels to which the transformed pixel corresponds. Pixel priorities are incorporated into the compression in a manner similar to that in [2]. Each pixel of the transformed image is left-shifted (multiplied by two) a number of times equal to the corresponding value in the transformed priority map. This modified transformed image is now compressed progressively one bit layer at a time, by a procedure similar to that used in the basic ICER. The packets produced are assigned priority values based on the bit layers for which they contain data. The output packets form a progressively coded “chain,” such that truncation of the chain at any point leaves a subset of packets that can be used to reconstruct the original image with a certain amount of distortion, which decreases monotonically with the number of packets retained. Each transformed pixel has a dynamic range which depends on the dynamic range of the original image and on the subband to which the pixel belongs. The process of modifying the transformed image by left-shifting pixels changes the bit positions for which the pixels are necessarily zero, and in addition these positions now can vary within a subband. Thus, to maintain compression efficiency, ROI ICER contains provisions for compressing only those bits which are allowed to be nonzero. The decompressor requires as input the same priority map supplied to the compressor. After decoding the (approximation to) the modified transformed image, the pixels of this image are right-shifted by the values in the transformed priority map, thereby undoing the opposite process which occurred during compression. The input images are color RGB images. Before compression, the images are transformed into the YCrCb domain, producing three new sub-images, one for each component, all referred to the same priority map. These images are compressed separately and produce separate chains of output packets with priority values. The priority values are suitable for comparing the priorities of packets from different components. C. BUFFER MANAGEMENT ALGORITHMS Our baseline buffer manager uses a simple form of double-valued prioritization. It determines all admissions and discards from the buffer according to the packet priorities computed by ROI ICER. However, the order of transmissions from the buffer is based on a simple first-in, first-out (FIFO) prioritization among all packets that survive the ROI-prioritized admission/discard process. Using a FIFO transmission priority insures that all transmissions will consist of truncated chains of packets as produced by ROI ICER. This eliminates the need to unshuffle the packets received on the ground, because the successive (truncated) packet chains can be used to reconstruct the source images in the same order in which they were acquired, but to different levels of distortion depending on how many packets from each chain survived the prioritized admission/discard process. The buffer management algorithm described here is the same with respect to buffer admissions and discards as the “on-line algorithm with buffer sorting” described in Section III.C, but that algorithm lso does full sorting by packet priorities to determine transmissions. Thus far, we have not noticed any large penalty for substituting the much simpler FIFO transmission protocol as long as we use the ROIcoded packet priorities to determine admissions and discards. But this tradeoff is a subject for further research. III. R ESEARCH ONFUTURE ALGORITHMS In parallel with developing our baseline system, we are also conducting research into various algorithmic approaches whereby the system performance might be improved or better understood. A. FEATURE COMPRESSION AND CLASSIFICATION FOR PRIORITIZATION FEEDBACK Priority should be given to information that has high science value, and in particular to information that is novel. This novelty should be determined with respect not only to a priori notions of what is to be expected, but also with respect to what has recently been transmitted. In other