Abstract
The robustness and accuracy of feature descriptors are two essential factors in the process of image registration. Existing feature descriptors can extract important image features, but it may be difficult to find enough correct correspondences for sophisticated images. And these feature descriptors often require domain expertise and human intervention. The aim of this paper is to utilise Genetic Programming (GP) to automatically evolve feature descriptors which are adaptive to various images including remote sensing images and optical images. In this paper, a novel GP-based method (GPFD) is proposed to extract feature vectors and evolve image descriptors for image registration without supervision. The proposed method designs a set of simple arithmetic operators and first-order statistics to construct feature descriptors in order to reduce noise interference. The performance of the proposed method is evaluated and compared against five methods including SIFT, SURF, RIFT, GLPM and GP. These results demonstrate that the feature descriptors evolved by GPFD are robust to complex geometric transformation, the illumination difference and noise.
Introduction
Image registration—the process of aligning two or more images of the same scene taken at different times, from different viewpoints, or by different sensors—is a fundamental task in computer vision. Applications range from medical imaging and remote sensing to autonomous navigation. Traditionally, this process relies on hand-crafted feature descriptors like SIFT (Scale-Invariant Feature Transform) or SURF (Speeded Up Robust Features).
While effective for standard photography, these traditional descriptors often fail when faced with "sophisticated" imagery. High levels of noise (such as speckle noise in radar data), extreme illumination changes, or multi-modal data (e.g., matching optical photos to Synthetic Aperture Radar (SAR) images) present significant challenges. These scenarios typically require expert manual tuning, which is inefficient and lacks scalability. This paper introduces GPFD (Genetic Programming of Feature Distance), an automated approach to evolving robust descriptors that adapt to complex image characteristics without human intervention.
The GPFD Methodology
The core innovation of GPFD is the use of Genetic Programming (GP) to automate the design of feature descriptors. Instead of a fixed, human-designed formula, GPFD "evolves" mathematical expressions that best distinguish between matching and non-matching image regions.
Evolutionary Framework
The algorithm maintains a population of mathematical "trees." Each tree represents a potential descriptor formula composed of:
- Function Set: Basic arithmetic operators (addition, subtraction) and a specialized "code" function. The code function acts as a rectifier, replacing negative values with zero and positive values with one, allowing the GP to evolve logic-like structures.
- Terminal Set (Robust Statistics): To ensure noise resistance, the method avoids raw pixel values. Instead, it uses first-order statistics calculated over small image windows: the 25th percentile, 75th percentile, mid-value, and standard deviation. Percentiles are notably more stable than means or extremes in the presence of noise.
Training Process
The evolution is guided by a fitness function based on the $\chi^2$ distance, which measures how well a candidate descriptor separates true matches from false ones. Over successive generations, the best-performing formulas undergo crossover and mutation to produce an optimal descriptor.
Application and Testing Pipeline
Once the optimal descriptor is evolved, it is integrated into a registration pipeline. The system first uses SIFT with a relaxed threshold to find a broad set of potential matches. The evolved GPFD descriptor then filters these candidates to identify true correspondences, followed by a final RANSAC (Random Sample Consensus) step for geometric refinement.
Experimental Results
The GPFD method was evaluated across seven diverse datasets, including Multi-View Stereo, aerial imagery (CIAP), city buildings (ZuBuD), and highly noisy SAR-to-Optical pairs.
Accuracy and Robustness
GPFD consistently outperformed established methods like SIFT, SURF, and RIFT.
- SAR-to-Optical Registration: In this challenging task, GPFD achieved a Root-Mean-Square Error (RMSE) of 0.5512, the lowest among all tested methods.
- Noise Resistance: In SAR imagery plagued by speckle noise, GPFD identified 72 correct matches with perfect precision (1.0). In contrast, standard SIFT found only 6 matches, failing to provide a reliable registration.
- Recall Rate: GPFD demonstrated a significantly higher recall rate in complex scenarios, often doubling or tripling the performance of specialized methods like RIFT or GLPM.
Discussion and Conclusion
The success of GPFD lies in its ability to discover mathematical relationships that are "impervious" to noise by leveraging stable statistical terminals like percentiles. By automating the descriptor design process, the method removes the need for domain expertise while achieving superior robustness across varied image types.
Limitations and Future Work: The primary trade-off for this robustness is computational cost; the evolutionary search is more time-consuming than using pre-defined descriptors. Future research will focus on optimizing the construction of training sets to improve efficiency and exploring higher-order image statistics to handle even more extreme multi-sensor registration tasks.
Summary updated