Haar-like、HoG 、LBP 三种描述方法在目标识别中的优劣


Haar-like的优势在于能更好的描述明暗变化,因此用于检测正面的人脸

HoG的优势在于能更好的描述形状,在行人识别方面有很好的效果

LBP比haar快很多倍,但是提取的准确率会低(10-20% 取决于训练对象)如果是嵌入式或者移动端的开发,推荐使用LBP。

这也解释了为什么haar应用于人的正面检测要明显好于应用于侧脸检测:正脸由于鼻子等凸起的存在,使得脸上的光影变化十分明显。而侧脸侧脸最重要的特征是形状和轮廓。 所以用HoG描述符检测侧脸更加有效。

参考原文:

https://www.quora.com/Why-are-HOG-features-more-accurate-than-Haar-features-in-pedestrian-detection 

It’s important to look at the most prominent feature of pedestrians. There can be more than one prominent feature but the defining feature of a typical pedestrian is the outline, the legs and head shape. Hence the detection method that best captures or describes the pedestrian outline will ultimately solve the pedestrian detection problem more accurately. HoG features are capable of capturing the pedestrian or object outline/shape better than Haar features. On the other hand, simple Haar-like features can detect regions brighter or darker than their immediate surrounding region better than HoG features. In short HoG features can describe shape better than Haar features and Haar features can  describe shading better than HoG features.That is also why Haar features are good at detecting frontal faces and not so good for detecting profile faces. This is because the frontal face has features such as the nose bridge which is brighter than the surrounding face region. But the profile face most prominent feature is it’s outline or shape, hence HoG would perform better for profile faces. HoG and Haar-like features are complementary features, hence combining them might even result in better performance. HoG features are good at describing object shape hence good for pedestrian detection. Whereas Haar features are good at describing object shading hence good for frontal face detection.