| Popis: |
This paper presents a novel pattern-driven image compression technique for exploring the synergy between content-based analysis and compression. Within the pattern-driven paradigm, image data are considered as relational and classifiable entities, which are low-level visual patterns including: (1) flat or homogeneous patterns (HP), (2) structural lines, curves and boundaries indicating the intensity/structural discontinuities (SD), and (3) complex/composite patterns (CP). Concisely, the pattern-based image model f can be loosely defined as . With such a pattern-driven model, efficient compression can be achieved by designing and developing novel ways of modeling and encoding disparate low-level visual patterns. For a given image, disparate low-level visual patterns are automatically separated, modeled, and encoded by: (1) A geometric filter, captures and models the predominant structural lines/curves (SD), (2) A linear filter, breaks various image regions into variable size triangular tiles and applies linear regression techniques to efficiently model tiles with HP, and (3) non-linear filters, model CP content such as cluttered background and textures, using advanced learning and transform based coding techniques. After all the filtering pipelines, adaptive entropy encoding schemes are employed to further reduce the bit cost by eliminating the statistical redundancies exhibited by all the features extracted and modeling parameters. The feasibility and efficiency of the proposed technique were corroborated by quantitative experiments and comparisons with the existing compression standards JPEG and JPEG2000. Our results show that the newly devised technique brings evident advantages including better support for compressed-domain analysis and more satisfactory subjective quality, mainly due to: the separation of the visual patterns allows the customization of the encoding schemes to maximize the coding efficiency using a compact set of extracted features and parameters. Furthermore, depending on the given application, different features can be prioritized and configured to match the data characteristics and the visual patterns therein for performance optimization. Since different patterns are segmented and modeled explicitly during the compression process, the proposed technique holds a great potential for achieving a good synergy between compression and compressed-domain analysis. For example, when applied to high-fidelity Ortho-imagery in GIS, city planning, and map updating applications, the structural information explicitly extracted matches the vector features such as road networks and regional boundaries. This correlation could be substantial for utilizing a new compression format with better support for compressed-domain analysis. Future research will further improve the compression efficiency and use specific applications to quantitatively demonstrate the synergy between compression and compressed-domain analysis within the pattern-driven framework. |
