研究目的
To propose a new descriptor for pattern recognition in gray-scale images that is invariant to nonuniform illumination, based on a biologically plausible local energy model of the visual system, and to compare its performance with known descriptors under various distortions.
研究成果
The proposed descriptor, based on monogenic signal and phase congruency, is effective for pattern recognition in images with nonuniform illumination, noise, and minor geometric distortions, outperforming SIFT and SURF in computer simulations. It is biologically inspired and invariant to key distortions, with potential for rapid implementation using parallel technologies.
研究不足
The experiments are based on computer simulations with synthesized distortions; real-world image conditions may vary. The method is local in nature, which might limit its applicability to global image features. Implementation speed could be optimized with parallel programming, but this was not fully explored in the paper.
1：Experimental Design and Method Selection:
The methodology involves designing a feature descriptor using a local energy model, monogenic signal framework, and a modified histogram of oriented gradients algorithm based on phase congruency. Theoretical models include the local energy function, phase congruency function, monogenic signal definition, and Poisson representation for scale invariance.
2：Sample Selection and Data Sources:
Test grayscale images containing different objects distorted by geometric transformations (rotation, scaling, contrast changes), additive noise, and nonuniform illumination modeled using the Lambert model. Over six thousand synthesized images from three different scenes were used.
3：List of Experimental Equipment and Materials:
Computer for simulation; no specific hardware mentioned. Software or algorithms include the proposed MHOG descriptor, SIFT, and SURF for comparison.
4：Experimental Procedures and Operational Workflow:
Key points are detected using a modified Harris corner detector applied to monogenic signal components. A local neighborhood around each point is constructed, smoothed with a Gaussian window, divided into regions, and a Histogram of Oriented Phase Congruency (HOPC) is computed. Performance is evaluated based on the normalized number of correct matching events (matching score).
5：Data Analysis Methods:
Comparison of the proposed descriptor (MHOG) with SIFT and SURF using the matching score criterion under various conditions of nonuniform illumination, noise, and contrast changes. Results are presented graphically.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
