Abstract:This paper introduces the new and powerful image patch descriptor based on second order image statistics/derivatives. Here, the image patch is treated as a 3D surface with intensity being the 3rd dimension. The considered 3D surface has a rich set of second order features/statistics such as ridges, valleys, cliffs and so on, that can be easily captured by using the difference of rotating semi Gaussian filters. The originality of this method is based on successfully combining the response of the directional filters with that of the Difference of Gaussian (DOG) approach. The obtained descriptor shows a good discriminative power when dealing with the variations in illumination, scale, rotation, blur, viewpoint and compression. The experiments on image matching, demonstrates the advantage of the obtained descriptor when compared to its first order counterparts such as SIFT, DAISY, GLOH, GIST and LIDRIC.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to develop a new image patch descriptor (RSD - DOG) based on the second - order derivative of the image in order to improve the robustness and discrimination ability in image matching tasks. Specifically, the author hopes to provide better performance than existing first - order derivative - based methods (such as SIFT, DAISY, GLOH, GIST and LIDRIC) when facing common image transformations such as illumination change, scale change, rotation, blurring, view - point change and compression by using the second - order statistical characteristics of the image (such as ridges, valleys, etc.).
### Main problems and goals
1. **Improve the robustness of image feature descriptors**:
- Existing first - order derivative - based image descriptors (such as SIFT, DAISY, etc.) perform poorly when dealing with problems such as illumination change, scale change, rotation, blurring, view - point change and compression.
- This research aims to improve the robustness to these changes by introducing a second - order derivative - based descriptor.
2. **Utilize second - order statistical characteristics**:
- An image can be regarded as a three - dimensional surface, where intensity is the third dimension. This representation method allows the extraction of rich second - order features, such as ridges, valleys, cliffs, etc.
- The author proposes to use the difference of rotated semi - Gaussian filters (DHSF) to capture these second - order features and combine it with the traditional difference of Gaussians (DOG) method.
3. **Reduce the dimension of the descriptor**:
- Many descriptors based on second - order statistical characteristics usually have a high dimension, which will increase the computational complexity and storage requirements.
- The goal of the RSD - DOG descriptor is to maintain a low dimension (256 - dimensional) while still being able to effectively capture second - order features.
4. **Verify the effectiveness of the new descriptor**:
- The author verifies the performance of the RSD - DOG descriptor under different image transformations through a series of experiments, especially in the case of complex illumination changes, demonstrating its superiority.
### Formula presentation
In order to accurately describe the application of second - order derivatives and filters, some key formulas are used in the paper:
- **Rotational smoothing filter**:
\[
G(\mu, \lambda)(x, y, \theta)=C\cdot H\left(R_{\theta}\begin{pmatrix}x\\y\end{pmatrix}\right)\exp\left(-\begin{pmatrix}x&y\end{pmatrix}R_{\theta}^{-1}\begin{pmatrix}\frac{1}{2\mu^{2}}&0\\0&\frac{1}{2\lambda^{2}}\end{pmatrix}R_{\theta}\begin{pmatrix}x\\y\end{pmatrix}\right)
\]
where \(C\) is the normalization coefficient, \(R_{\theta}\) is the rotation matrix, \(x\) and \(y\) are pixel coordinates, and \(\mu\) and \(\lambda\) are the standard deviations of the Gaussian filter.
- **Directional signal \(D(x, y, \theta)\)**:
\[
D(x, y, \theta)=G(\mu, \lambda_{1})(x, y, \theta)-G(\mu, \lambda_{2})(x, y, \theta)
\]
where \(\lambda_{1}\) and \(\lambda_{2}\) are the standard deviations of two Gaussian filters with different widths.
- **Average angle and magnitude**:
\[
\begin{cases}
\eta_{1}(x, y)=\frac{\theta_{M1}+\theta_{M2}}{2}\\
\eta_{2}(x, y)=\frac{\theta_{m1}+\theta_{m2}}{2}\\
\delta_{1}=\frac{\left\|\ D(x, y, \theta_{M1})\right\|+\left\|\ D(x, y
\end{cases}
\]