Study On Image Histogram Feature And Application

Because of small computational cost and good invariance under transformation, rotation, scale, histogram technique is widely used in lots of image processing areas, especially in threshold segmentation, image retrieval based on color feature and image classification.Image segmentation is the basis of image recognition. We can use image segmen-tation to divide an image into two parts, background area and foreground area, so that we can avoid searching the whole area of the image when in image recognition, and great-ly improve accuracy rate of the result. Since threshold segmentation method based on grey image histogram is simple in computation, it is very suitable for grey image with simple scene, especially for remote sensing image.Image retrieval is a very important and challenging research topic nowadays. Color feature plays an important role in image retrieval, because it is very easy to be under-stood and has low computational cost. Early image retrieval algorithms mainly employ color feature, especially color histogram.Image classification task is to automatically process a group of images, and finally output every images’categories. Image classification would have a promising future and is hot and hard problem of computer vision. Among so many image classification algorithms, these algorithms based on bag-of-words model are the most classic and ef-fective. Image classification algorithm based on bag-of-words model has three steps. Firstly, we extract color, shape and other kinds of features and produce a visual "code-book". Secondly, we convert the image to visual words, and construct a visual words histogram to represent the image. Finally, we use visual words histogram as feature vector and employ classifier, such as SVM, to classify.There exists some drawbacks in existing threshold segmentation algorithms, such as easy to be interfered by noise and not robust enough. So, it is essential to propose a segmentation algorithm based on grey histogram which could avoid being interfered by noise and is robust. Traditional color histogram methods exist many drawbacks, such as high dimensions, illumination effects, insufficient of conveying correlation of similar colors, and losing spatial information, so it is necessary to propose a new color histogram method with systematically considering all of these factors. The key of the bag-of-words model is to extract effective visual words histogram to describe image, while recent existing bag-of-words methods are still not good enough in feature selec-tion and word frequency statistic. So we need to propose a new bag-of-words based method to solve these problems.In this paper, we have three aspects of research:threshold segmentation based on grey image histogram, color spatial histogram and image classification base on bag-of-words model. The main works are as follows:1.1-D threshold segmentation algorithm has poor adaptability and poor noise im-munity, while2-D threshold segmentation has high computational complexity. To solve these problems, we propose a novel threshold segmentation algorithm. The whole method contains three steps:Firstly, we construct a1-D histogram by using each pixel’s grey value and its surrounding area’s mean grey value. Then, we propose a new threshold selection algorithm by merging three classic segmen-tation algorithms. Finally, we use the selected threshold to instruct segmentation. The experimental results show that our algorithm has strong adaptability, stable and robust noise immunity when comparing to1-D threshold segmentation al-gorithm and a wider range of noise adaptability, lower computation complexity when comparing to2-D Otsu algorithm.2. Traditional color distance is usually measured by simple Euclidean distance for-mula, but in HSV color space, this kind of simple distance measurement is not suitable, because each component’s contribution degree to color is different:H component is highest, and V component is lowest. Hence, we can not employ simple color distance measurement in HSV color space. To solve this problem, we propose a parameterized HSV color space distance formula, by using parame-ter to distinguish different components. We manually label data, and utilize pair-wise based learning method to train the labelled data and get the parameters of predefined distance formula. Finally, we get a parameterized HSV color space distance formula.3. There are four problems on color histograms:the dimension of color histograms is too large; color histograms are easy affected by the change of illumination; sim-ilar colors are not relevant; and color histograms lack spatial information. A nov-el algorithm of cluster-based spatial color histograms is proposed to solve these problems. Firstly, this method runs k-means cluster on the source image, and constructs spatial color histograms on the image created by k-means cluster. We train the parameters of HSV-space distance formula, and then propose a histogram matching algorithm based on the most similar colors iteration. Our cluster-based spatial color histograms feature is used in image retrieval experiments, compared with other color histograms based image retrieval methods, precision and recal-1are improved. The results show that our method can describe subjective and objective image color features very well, and has stronger adaptation and robust-ness.4. Bag-of-words is classical method for image classification. The core problem is what visual words to select and how to construct the visual words histogram. In this paper, we propose a visual attention based bag-of-words(VABOW) model for image classification task. The VABOW model utilizes visual attention method to generate a saliency map, and uses the saliency map as a weighted matrix to instruct to generate the "codebook" and construct the visual words histogram. On the other hand, the VABOW model uses L1regularization logistic regression method to select the most relevant and most efficient features. We compare our approach with existing bag-of-words based methods on two datasets, and the re-sult shows that our VABOW model outperforms the state-of-the-art method for image classification.