We define localized content-based image retrieval as a CBIR task where the user is only interested in a portion of the image, and the rest of the image is irrelevant. In this paper we present a localized CBIR system, Accio, that uses labeled images in conjunction with a multiple-instance learning algorithm to first identify the desired object and weight the features accordingly, and then to rank images in the database using a similarity measure that is based upon only the relevant portions of the image. A challenge for localized CBIR is how to represent the image to capture the content. We present and compare two novel image representations, which extend traditional segmentation-based and salient point-based techniques respectively, to capture content in a localized CBIR setting.

Image segmentation is a fundamental step in many computer vision applications. Generally, the choice of a segmentation algorithm, or parameterization of a given algorithm, is selected at the application level and fixed for all images within that application. Our goal is to create a stand-alone method to evaluate segmentation quality. Stand-alone methods have the advantage that they do not require a manually-segmented reference image for comparison, and can therefore be used for real-time evaluation. Current stand-alone evaluation methods often work well for some types of images, but poorly for others. We propose a meta-evaluation method in which any set of base evaluation methods are combined by a machine learning algorithm that coalesces their evaluations based on a learned weighting function, which depends upon the image to be segmented. The training data used by the machine learning algorithm can be labeled by a human, based on similarity to a human-generated reference segmentation, or based upon system-level performance. Experimental results demonstrate that our method performs better than the existing stand-alone segmentation evaluation methods.

In the liquid architecture project, we are exploring ways in which architectural flexibility can be exploited to improve the execution properties of individual applications. Here, we report on successes we have had to date in this area, and present our vision of where this research should proceed into the future.

In this paper, we present two new methods for efficient rate control and entropy coding in lossy image compression using JPEG-2000. These two methods enable significant improvements in computation complexity and power consumption over the traditional JPEG-2000 algorithms. First, we propose a greedy heap-based rate-control algorithm (GHRaC), which achieves efficient postcompression rate control by implementing a greedy marginal analysis method using the heap sort algorithm. Second, we propose an integrated rate-control and entropy-coding (IREC) algorithm that reduces the computation complexity of entropy coding by selectively entropy coding only the image data that is likely to be included in the final bitstream, as opposed to entropy coding all image data. Together, these two methods enable significant savings in computation time and power consumption. For example, the GHRaC method demonstrates 16 /spl times/ speedup for rate control when encoding the Lena color image using a target compression ratio of 128:1, one quality layer, and code blocks of 32 /spl times/ 32 pixels. The IREC method expands upon GHRaC to perform entropy coding in conjunction with rate control. Using an enhanced version of IREC, these two methods jointly achieve a speedup in execution time of 14 /spl times/ over traditional rate control and entropy coding, which first entropy codes all image coefficients and then separately performs postcompression rate control using the generalized Lagrange multiplier method to select which data are included in the final bitstream. Both theoretical analysis and empirical results are presented in validating the advantages of the proposed methods.

This paper presents a hierarchical image representation that preserves spatial relationships, and an image segmentation algorithm that efficiently constructs this representation. The overall structure of the hierarchical representation assumes a tree-like organization, in which a node at a higher level represents of a group of nodes from the next lower level. The root node of the representation is the full image, and the leaf nodes are K /spl times/ K blocks of pixels (usually 4 /spl times/ 4 or 8 /spl times/ 8), called blobs. The segmentation algorithm builds this hierarchical representation using an efficient clustering algorithm for grouping nodes progressively into larger objects. While performing the union of nodes at one level into a single parent node at the next higher level of the tree, nodes are grouped together according to similarities between their feature vectors, which include such features as color information, orientation, texture, size, energy, and neighbor information. The computation complexity of the hierarchical image segmentation algorithm is O(MN + MN/K/sup 2/ log(MN/K/sup 2/)), where M, N are the height and width of the original image. Experimental results show that this algorithm can segment images and extract objects effectively.