摘 要：目的 为克服单一颜色特征易受光照变化影响，以及图像的空间结构特征对目标形变较为敏感等问题，提出一种结合颜色属性的分层结构直方图。方法 首先，鉴于使用像素灰度值对图像进行分层易受光照变化影响，本文基于颜色属性对图像进行分层，即将输入的彩色图像从RGB空间映射到颜色属性空间，得到11种概率分层图；之后，将图像中的每一个像素只投影到其概率值最大的分层中，使得各分层之间像素的交集为空，并集为整幅图像；对处理后的每一个分层，通过定义的结构图元来统计其像素分布情况，得到每一分层的空间分布信息；最后，将每一分层的像素空间分布信息串联作为输入图像的分层结构直方图，以此来表征图像。结果 为证明本文特征的有效性，将该特征用于图像匹配和视觉跟踪，与参考特征相比，利用本文特征进行图像匹配时，峰值旁瓣比均值提升1.3479；将本文特征用于视觉跟踪时，采用粒子滤波作为跟踪框架，成功率相对上升4%，精度相对上升4.6%。结论 该特征将图像的颜色特征和空间结构信息结合，有效的解决了单一特征分辨性较差问题，与参考特征相比该特征具有更强的分辨性和鲁棒性，因此本文特征可以更好地应用于图像处理应用中。
Hierarchical structure histogram combined with color name
Yue Chen Chen,HOU ZHI QIANG()
Abstract: Objective The extraction of image feature plays a very import role in the field of visual tracking and is an essential component for image matching and other image processing applications. At the same time, different image description methods can affect the performance of the algorithm directly. In recent years, domestic and foreign researchers have proposed numerous image features, which can be divided into two classes as follows: 1) deep features, based on deep learning, has obtained successful effect, but it needs a large number of data to train the model and has a great demand on the experimental platform, which has some restrictions on its applications to a great extent. 2) traditional manual features, which can be performed on any existing platform with simple and intuitive. Besides, it has fulfilled remarkable results in image processing too, including SIFT(Scale-invariant feature transform), HOG(Histogram of Oriented Gradient) and CN (Color Name), etc. So it is very essential and necessary to further study manual features. Nevertheless, it is difficult to improve the performance of the algorithm only depending on a single feature. In this paper, a hierarchical structure histogram combined with color name is proposed in order to overcome the problem that a single color feature is so susceptible to illumination changes that lead to poor robustness, and that the spatial structure of the image is sensitive to the deformation of the target which will pull down the distinguishability of feature. Method Aimed at the disadvantages that pixel gray value is susceptible to the layered image with the illumination change, an improved method that according to color name to layer image is proposed in this work. Firstly, it project original RGB color space to a more robust color space-color name(CN) space, and object are represented by a probabilistic 11 dimensional map, which means the input image is stratified to 11 layers based on color name. Then, each pixel in the image is projected into the layer with the highest probability, so that the pixels intersect with each other to be empty, and are integrated into the whole image. That is to say, each pixel can only be projected into one layer. Furthermore, for the every hierarchical image, it calculates the distribution of pixels on every dimension by counting the number of pixels in each square of structure image element, then gets the spatial distribution information of pixels. Finally, the pixel spatial information of each slice is connected in series as the hierarchical histogram to represent the image. Result In order to prove the validity and strong distinguishability of the proposed feature, two experiments have been carried out in this paper. The first one is image matching, whose strategy is that there are an extant model, then the position of the matched image is determined by exhaustively traversing the original image. The data set of image matching comes from PASCAL VOC2007(Visual Object Classes), which contain various classes of object(e.g., person, bird, car, dog and so on). When calling this data set, only the target pointed by the first Ground Truth of each image is used. The other one is visual tracking, whose tracking frame adopts particle filter and the number of particles is 200. This experiment is evaluated on 100 sequences of Object Tracking Benchmark(OTB100), which mainly contains 11 challenges(e.g., out-of-plane rotation, scale variation, illumination variation and so on) that may be encountered in object tracking. All the experiments in this paper are run on the windows platform and the development environment is MATLAB. We compare this feature against other 4 traditional manual features and the experimental results show that image matching based on this features can locate the target more accurately, distinguish similar targets well and the peak value of the target is the most obvious, in the meantime the mean value of peak-to-side lobe ratio increases by 1.3479. Moreover, object tracking based on it can reduce the center location error of tracking significantly and improve accuracy and success rate momentously, meanwhile the success rate rises by 4% and the accuracy goes up by 4.6%. Conclusion Hierarchical structure histogram based on pixel gray value is easily affected by illumination and target rotation changes. So, in this paper, we adopt a new method to layer image to gain robust feature. The proposed feature combines color name features with spatial information of pixels in image to improve the capability of feature to adapt to various scenes, such as illumination changes, deformation, and low resolution, which effectively enhances the poor discrimination of single feature and makes it more robust in image matching and object tracking, especially for the objects with the same color distribution but different spatial distribution of pixels. In addition, this feature remains the representation of traditional histogram features, which making computation and similarity measurement simpler. Compared with other 4 traditional manual features, this feature exhibits better image matching performance and visual tracking result in most cases. Therefore, this feature can be better applied to image processing applications. However, this feature in the research process remains deficiencies, which does not consider the problem of target scale change in the visual tracking process. In the follow-up work, we will further optimize the feature to obtain enhanced generalization capabilities in visual tracking and remain applicable to other image processing applications simultaneously.
Key words: image matching; visual tracking; color name; spatial information; hierarchical structure histogram