Abstract
This paper introduces a novel framework for 2D shape recognition that is inherently invariant to Euclidean transformations—rotation, translation, and scale. By combining Local Fourier-Mellin Descriptors (LFMD) with the Vector of Locally Aggregated Descriptors (VLAD), the authors demonstrate a system that achieves state-of-the-art accuracy while requiring significantly less training data than standard Convolutional Neural Networks (CNNs). The approach avoids the need for massive data augmentation by embedding geometric invariance directly into the descriptor mathematics.
Introduction
A fundamental challenge in computer vision is achieving Euclidean invariance: the ability to recognize an object regardless of its position, orientation, or size. While modern CNNs are the dominant solution, they typically lack mathematical invariance. Instead, they rely on "data augmentation," where thousands of transformed versions of an image are used during training to help the network memorize variations. This process is computationally expensive and data-inefficient.
Traditional mathematical transforms like the Fourier-Mellin transform offer built-in invariance but are highly sensitive to the choice of the coordinate origin. The authors address this by proposing a localized version of these transforms, creating a robust and efficient descriptor for shape and texture recognition.
Methodology
The core of the proposed system is the Local Fourier-Mellin Descriptor (LFMD). Unlike global transforms, LFMDs are computed at every pixel within a local neighborhood.
Log-Polar Harmonic Gratings
The system utilizes log-polar coordinate mapping to transform rotations and scales into simple translations. In this coordinate system, a rotation in the image plane becomes a shift along the angular axis, and a change in scale becomes a shift along the logarithmic radial axis.
Localization and Stability
To ensure the filters only capture local information, the basis functions are multiplied by a polynomial envelope. The authors found that a parameter $\alpha < -1$ is necessary to ensure the filters are sufficiently localized to capture fine details. Additionally, a "hole strategy" is employed to zero out the unstable center of the log-polar coordinate system, preventing mathematical aliasing.
LFMD-VLAD Integration
To achieve translation invariance, the local descriptors are pooled using the Vector of Locally Aggregated Descriptors (VLAD) method. This creates a global histogram of local features. Because histograms are independent of the spatial location of the descriptors, the final representation is fully invariant to the object's position in the frame.
Key Results and Data Efficiency
The system was evaluated against AlexNet on several benchmarks, including a modified MNIST dataset with random rotations and scaling.
Performance Comparison
On the rotated and scaled MNIST dataset, LFMD-VLAD achieved a 3.12% error rate using only the original, non-augmented images for training. In contrast, AlexNet required a dataset ten times larger (heavily augmented) to reach a comparable error rate of approximately 6%.
Shape and Texture Robustness
On the Swedish Leaves dataset, the system reached 95.84% accuracy, matching specialized shape recognition algorithms. It also outperformed CNNs on the KTH-TIPS texture database when training data was extremely limited (e.g., only 5 images per category), proving that hard-wired invariance is a powerful alternative to learned invariance in low-data regimes.
Conclusion
The LFMD-VLAD framework demonstrates that mathematical invariance can significantly reduce the "data hunger" of computer vision systems. By preserving essential visual information—as evidenced by successful image reconstruction experiments—while discarding phase information for invariance, the descriptors provide a robust foundation for efficient recognition. Future work will focus on extending these descriptors to handle cluttered backgrounds and developing hierarchical deep networks with built-in Euclidean invariance.
Summary updated