Because the contrast value and the two-bin contrast histogram are used

contrast context histogram (CCH) has the characteristics of low dimension and efficient computation. The CCH contrast value is calculated between every pixel in a local region and the center pixel of the region making the CCH contrast values be sensitive to the intensity changing of the center pixel. Usually

the stability of the center intensity depends on the accuracy of the local feature detector position

which is susceptible to noise

image transformation and distortion

scale space discretization

and other factors. Furthermore contrast values in CCH are not constant illumination invariant because the intensity of the center point is not the mean intensity of the whole local region in most instances. As a result

CCH descriptors cannot get competitive performance when compared to the standard SIFT (scale invariant feature transform) descriptor. In addition

the SIFT descriptor has the characteristics of high dimension and low computational efficiency. In order to solve these problems

a fast and low-dimensional local descriptor is proposed

which is called mean normalized contrast context histogram (MN-CCH). MN-CCH firstly calculates the Gaussian weighted mean of the whole local region. The core of the Gaussian kernel coincides with the center of the local region. The local region is normalized by its weighted mean to obtain the region's normalized contrast values. Then

the local region is divided into 32 sub-regions in the log-polar coordinate system

and a two-bin positive-negative contrast histogram is built for each sub-region. To overcome the linear light changes

the 64-dimension MN-CCH vector is normalized to a unit vector. As the Gaussian weighted mean is used

MN-CCH contrast values are more stable for the feature location error when compared to the center point intensity used in CCH. In addition

the normalized contrast values are constant illumination invariant

which helps to improve the MN-CCH's performance. The image matching experiment in the Mikolajczyk image transformation dataset shows the proposed MN-CCH outperforms the CCH descriptor in most image transformations

especially in natural images and texture images. The image matching result also shows that the MN-CCH overcomes the low Recall in lower 1-Precision and noise-sensitive problems exist in the CCH descriptor. The proposed 64-dimension MN-CCH gets competitive performance when compared to 128-dimension SIFT descriptor in the image matching experiment. The retrieval test on the PCA-SIFT's retrieval dataset shows that MN-CCH gets the same retrieval accuracy as SIFT does

which is higher than the CCH descriptor. MN-CCH's descriptor building and matching time is indentical to the CCH as these two methods have the same dimension and complexity

which is more efficient than the SIFT descriptor as SIFT's tangential calculation in descriptor building is time-consuming and SIFT uses a higher descriptor dimension. MN-CCH outperforms the CCH descriptor in both image matching and retrieval tests as it overcomes CCH's feature location error sensitive and contrast values are not constant illumination invariant

drawbacks. The proposed method has higher descriptor building speed and lower descriptor dimension

but it still gets competitive performance when compared to the SIFT descriptor. As a result

MN-CCH is more suitable in applications requiring high computing and large storages resources

such as mobile robot real-time navigation

visual SLAM and so on.