EFFICIENT LEARNING OF MULTI-IMAGE STRUCTURE
Abstract
The ubiquity of data-driven approaches in computer vision has been fuelled by the availability of large-scale data, both in the form of publicly available datasets and actual field data, as well as cheap, high-end compute power. Data driven approaches have significantly increased the importance of efficient and effective representation of images. Rich data representations, efficient algorithms, and high performance computing environments permit us to derive meaningful inferences from the data and keep the processing tim es at tolerable levels. Mostrepresentations in use today are obtained by training systems in an end-to-end, bottom-up fashion. The work proposed here is on two themes, motivated by the needs of data driven visual computing. One of them is on network compression for efficient processing while the other is on the design of effective algorithms for extracting representations. On the first theme of efficient processing, this proposal is aimed at improving the computational efficiency of popular convolutional neural networks (CNNs), by shrinking the network to smaller sizes without significantly compromising on their performance.The goal is effective network size compression to reduce the large number of parameters present in CNNs, to make them run better in resource-constrained environments.On the second themeof efficient tools for extracting representations, development of a new approach to clustering is proposed. While clustering has been a topic of research for quite sometime, it has found extensive and increasing use in computer vision in recent years. However, there has been limited work on enhancing the representation power of clusters, andcomputational efficiency with which they are extracted to suit the changing needs of the data driven models. For instance, exploiting local structure of the spatial distribution of data points has not received enough attention for improving performance of clustering algorithms. This objective is formulated here as an energy minimization problem. The third part of the proposed work, still on the representation theme, integrates the aforementioned work, for providing an improved solution to the computer vision problem of discovering and delineating any recurring, a priori unknown objects in multiple images. The goal is toidentify and detect object (instead of low) level structures present across multipleimages. This problem, also known as co-segmentation, has received limited attention but is of increasing importance in the data driven mode of processing.The computer vision problems used here for testing the tools exploit structure inherent in images. The results of the proposed work on representation and processing extend beyond the particular applications chosen here to develop them. Since the objectives they serve are common across many computer vision problems, they may be seen as computer vision tools for broad use.To a limited extent, the proposed work also tests transferability of the proposed approaches to a different domain, that of detection of anomalous activities in computer usage data. Objects/events associated with secure computer operations, analogous to those in images but with different definitions, are looked for, to take advantage of the high-precision of the aforementioned models for predicting the occurrence of such events. This project is anticipated to deliver state-of-the-art methods for the CNN compression, data clustering, image co-segmentation, and attack detection. These algorithms would be communicated as scientific publications to reputable journals or conferences.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 17, 2020
- Source ID
- N000142012444
Entities
People
- Narendra Ahuja
Organizations
- Office of Naval Research
- United States Navy
- University of Illinois Urbana–Champaign