Abstract:Local binary pattern (LBP) as a kind of local feature has shown its simplicity, easy implementation and strong discriminating power in image recognition. Although some LBP variants are specifically investigated for color image recognition, the color information of images is not adequately considered and the curse of dimensionality in classification is easily caused in these methods. In this paper, a color-related local binary pattern (cLBP) which learns the dominant patterns from the decoded LBP is proposed for color images recognition. This paper first proposes a relative similarity space (RSS) that represents the color similarity between image channels for describing a color image. Then, the decoded LBP which can mine the correlation information between the LBP feature maps correspond to each color channel of RSS traditional RGB spaces, is employed for feature extraction. Finally, a feature learning strategy is employed to learn the dominant color-related patterns for reducing the dimension of feature vector and further improving the discriminatively of features. The theoretic analysis show that the proposed RSS can provide more discriminative information, and has higher noise robustness as well as higher illumination variation robustness than traditional RGB space. Experimental results on four groups, totally twelve public color image datasets show that the proposed method outperforms most of the LBP variants for color image recognition in terms of dimension of features, recognition accuracy under noise-free, noisy and illumination variation conditions.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is that in color image recognition, the existing Local Binary Pattern (LBP) variants cannot fully consider color information and are likely to lead to an increase in the dimension of feature vectors and a decline in classification performance. Specifically, although some LBP variants have been specifically studied for color image recognition, they do not fully utilize the color information of the image and are easily affected by the curse of dimensionality during the classification process. Therefore, this paper proposes a new color - related Local Binary Pattern (cLBP), aiming to improve the performance of color image recognition by learning dominant patterns from decoded LBPs, while reducing the dimension of feature vectors, thereby enhancing the discriminative ability of features.
### Main contributions
1. **Relative Similarity Space (RSS)**: A Relative Similarity Space (RSS) is proposed to represent the color similarity between each channel of a color image. RSS can better describe the color information in color images and has higher robustness to noise and illumination changes.
2. **LBP decoding**: Based on RSS and the traditional RGB color space, the traditional LBP operation is used to extract the feature maps of each channel, and then the joint information between different color channels is captured by decoding these LBP feature maps.
3. **Feature learning strategy**: A feature learning strategy is proposed to learn the dominant color - related patterns from the decoded LBP feature vectors, so as to reduce the dimension of feature vectors and further improve the discriminative ability of features.
### Experimental results
The experimental results show that the proposed cLBP method outperforms most existing LBP variants on a total of 12 publicly available color image datasets in four groups, whether under no - noise, noisy or illumination - changing conditions, especially in terms of feature vector dimension and recognition accuracy.
### Formula summary
- **Relative similarity**:
\[
RS=\frac{\vert p - q\vert}{p+\xi}
\]
where \(p\) and \(q\) are two positive numbers (the value range of RGB channels is [0, 255]), and \(\xi\) is a small value used to avoid a zero denominator.
- **Relative similarity calculation**:
\[
RSR(x,y)=\frac{\vert R(x,y)-G(x,y)\vert+\vert R(x,y)-B(x,y)\vert}{R(x,y)+\xi}
\]
\[
RSG(x,y)=\frac{\vert G(x,y)-R(x,y)\vert+\vert G(x,y)-B(x,y)\vert}{G(x,y)+\xi}
\]
\[
RSB(x,y)=\frac{\vert B(x,y)-R(x,y)\vert+\vert B(x,y)-G(x,y)\vert}{B(x,y)+\xi}
\]
- **Traditional LBP calculation**:
\[
LBP_{n,R,P}(x,y)=\sum_{m = 0}^{P - 1}S(V_{m,n,R,P}(x,y)-V_n(x,y))\times2^m
\]
where
\[
S(x)=\begin{cases}
1, &\text{if }x\geq0\\
0, &\text{if }x < 0
\end{cases}
\]
- **Decoded LBP feature value**:
\[
dM_m(x,y)=\sum_{n = 1}^62^{(6 - n)}\times S(V_{m,n,R,P}(x,y)-V_n(x,y))
\]
- **Decoded LBP feature map**:
\[
DLBP_c(x,y)=\sum_{m = 1}^P DLBP_m^c(x,y)\times2^{m - 1}
\]
where
\[
DLBP_m^c(x,y)=\begin{cases}
1, &\text{if }