【作者】 张晓红；

【导师】 刘岩；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2015， 硕士

【摘要】 在目标识别领域,人们已经取得了不错的成果,目标识别技术已应用到各行各业都可以看到它的应用,比如在生物领域,军事领域,农业领域等。随着研究的深入,人们开始将注意力转移到动物的检测上。受人脸检测成功的启发,考虑将人脸检测的技术应用到动物检测上,本文开展了对猫脸进行识别的工作。本文采取有粗到精的思想对猫脸进行检测,工作分为两步,第一步是粗检测,获得最有可能包含猫脸的候选区域,第二部分是对第一步的基础上应用可变形模板对猫脸进行精检测,得到最终结果。第一个工作充分利用了猫所具有的丰富的颜色和纹理信息。颜色特征是一种比较常见的图像底层特征,而且猫具有丰富的颜色信息。为了使得objectness方法生成的候选边界框更加针对猫,本文将图像转换到HSV颜色空间后,应用高斯混模型对颜色空间进行分割,从实验中可以看到猫的大致区域能够被分割出来。获得分割的结果之后,在其上应用objectness方法来产生候选矩形,由于加入了颜色的先验信息,使得选用少量的矩形框就能够比较好地标定猫所在的区域。纹理特征是一种比较常见的图像特征,LBP特征是描述纹理最有效的特征之一。在猫脸识别中,为了提取丰富的纹理信息,同时获得特征的局部信息,本文在提取的LBP特征上应用空间金字塔SPM,获得特征的空间几何信息。而为了更好地捕捉图像不同空间区域之间的纹理变化,本文将Haar-like思想应用到SPM中,提取空间区域之间的一阶信息来反映区域间的纹理的变化。在基于Microsoft Cat Dataset进行处理所得到的猫脸数据集上的实验结果验证了所提出方法的有效性。在该数据集上提取完特征之后,我们训练Adaboost级联分类器,进而获得猫脸的候选区域,从实验中可以看出,我们能够得到较为准确的候选区域。结合第一步的两个工作的结果,我们获得了关于猫脸的粗检测区域。在第二步中,为了能够在候选区域上精确地检测出猫脸,本文训练得到猫脸的可变形部件模型,在得到的包含猫脸的候选矩形上,应用DPM进行检测。实验证明这种基于区域的检测要好于整幅图像的检测效果,因而获得了比较好了检测效果。更多还原

【Abstract】 In the field of object recognition, researchers have achieved good results.The technology of object recognition has been applied to all walks of life, such as in the biological field, military areas, agriculture and so on. Recently, people pay more attention to the animals detection. Inspired by the success of face detection, we consider apply its technology to the animal detection.This paper does works around cat face detection.We adopts the thought of coarse to fine to detect cat face, which include two steps, the first to select candidate rectangles that most likely contain the cat face, the second to apply the deformable parts model to the former results to find the accurate position.The first job utilize the rich color and texture information of cat. The color feature is one of the most common low-level features and cat has rich color information. In order to make the objectness method generate candidate bounding boxes to favor cats more, we apply the Gaussian Mixture Model to the segmentation of the color space, after conversion of the image from RGB to the HSV color space. the experiment results show that the position of the cat in the image can be divided roughly. Based on the results, we run objectness method to generate bounding box. With the joined priori information, we can get more accurate candidate areas while considering fewer bounding boxes.As the most common image features in computer vision, texture can be represented effectively using LBP. In order to make full use of the rich texture information and obtain the local feature information, we apply the Spatial Pyramid Model on the extracted LBP feature to get the spatial information of the feature.Meanwhile, based on the idea of Haar-like feature, we apply the rectangle model on the SPM to capture the variations of features of different regions in the image. We conduct the experiments on the processed Microsoft Cat Dataset and the results validate the effectiveness of the proposed method. After extracting features of the dataset,, we train the cascaded Adaboost classifier and then get the candidate regions.Combining the two results in the first step, we obtain the position where the cat locate roughly. In order to accurately detect object in the second step, we train the deformable part model of cat face. Then we apply the DPM to detect cat face on the candidate rectangles. We conduct the experiment on the Microsoft Cat Dataset and get a better detection rate about cat face than other methods. And the results prove that it’s better to detect cat face on the local area than on the whole image.更多还原