Most of the existing vehicle logo recognition (VLR) approaches simply combine traditional image features with different classifiers

such as template matching

HOG(histogram of oriented gradient)/SVM(support vector machine)

HOG/KNN(k-nearest neighbor)

SIFT(scale-invariant feature transform)/SVM

and radial Chebyshev moments/SVM

but they do not consider structure features of vehicle logo image. Vehicle logo images have distinct structure information

which is the significant clue for VLR. Considering both their gray and structure features

this study proposes a new VLR approach based on pixel-pair features that are randomly sampled from foreground-background skeleton areas. First

a standard vehicle logo image is partitioned into foreground and background areas. Second

the skeleton areas of both foreground and background are extracted

which contracts foreground and background areas

because pixels in skeleton areas are more stable. Third

pixels are randomly selected from the two skeleton areas

forming pixel pairs by cross-matching that can obtain more pixel pairs. Then

through pixel-pair validation

accepted pixels are selected to express vehicle logo features. Features express similarity of local areas centered on both pixels

which reflect the gray relationship of logo and background areas after the imaging process. Pixel-pair validation has two crucial rules. One is that the pixel pair in a different sample of the same class has the same similarity relationship; and the other is that the pixel pair in a different sample of a different class has a different similarity relationship. Last

the connection relationship is introduced during classification. Through the connection relationship

independent pixel pairs

which contribute to vehicle logo recognition

can be aggregated. Experimental data derived from 19 044 vehicle logo images are captured by surveillance cameras in the real world. This image group is called Test Set 1. To test the performance of different methods under weak illumination

2 326 vehicle logo images with weak illumination are selected in Test Set 1

which is now called Test Set 2. Experiment shows that compared with recognition results based on other image features

the method based on the proposed feature can achieve higher recognition performance. Test Set 1 has a 95.7% recognition rate. Under weak illumination

our method attains an 87.2% recognition rate on Test Set 2. Considering both gray and structure features of vehicle logo images

we propose a new VLR approach based on pixel-pair features

which are randomly sampled from foreground-background skeleton areas. Pixel-pair feature extraction process fully considers the logo shape

which makes this feature distinctive and exclusive. Experiments show that our method attains a higher recognition rate compared with traditional VLR methods. Our method has shown strong robustness especially under weak illumination.