面部识别算法是如何工作的？

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"过去十年，深度学习领域出现了许多先进的新算法和突破性的研究，并且引入了新的计算机视觉算法。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这一切始于2012年的AlexNet。AlexNet是一个深度（卷积）神经网络，它在ImageNet数据集（拥有超过1400万张图片的数据集）上取得了很高的准确率。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"人类是如何识别人脸的？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"也许，人类大脑中的神经元首先识别场景中的人脸（从人的体形和背景），然后提取面部特征，并通过这些特征对人进行分类。我们已经在一个无限大的数据集和"},{"type":"link","attrs":{"href":"https:\/\/www.engati.com\/glossary\/neural-networks","title":null,"type":null},"content":[{"type":"text","text":"神经网络"}]},{"type":"text","text":"上进行了训练。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"机器中的"},{"type":"link","attrs":{"href":"https:\/\/www.engati.com\/glossary\/facial-recognition","title":null,"type":null},"content":[{"type":"text","text":"面部识别"}]},{"type":"text","text":"是以同样的方式实现的。首先，我们采用面部检测算法来检测场景中的人脸，然后从检测到的人脸中提取面部特征，最后使用算法对人进行分类。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/39\/11\/3983b31c3e2edf64f66352186d8b9311.png","alt":null,"title":"面部识别系统的工作流","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"1. 人脸检测"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"人脸检测是物体检测的一个特化版本，特别之处在于，它只检测一种物体，即人脸。就像计算机科学里需要权衡时间和空间，机器学习算法也需要在推理速度和准确性之间进行权衡。现在有很多物体检测算法，不同算法对速度和准确性的取舍有所不同。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"本文评估了如下几个最先进的物体检测算法："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"OpenCV（Haar-Cascade）"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"MTCNN"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"YoloV3和Yolo-Tiny"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"SSD"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"BlazeFace"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"ShuffleNet和Faceboxes"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"为了构建一个强大的人脸检测系统，我们需要准确且快速的算法，以满足在GPU以及移动设备上实时运行的需要。"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"准确度"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在流媒体视频的实时推理中，人们的面部可能有不同的姿势、遮挡和照明效果。因此，算法能在不同的光照条件和不同姿态下精确检测人脸非常重要。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":""}]},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/ac\/5b\/ac93bc1bc84bacda8b1426402e91d35b.png","alt":null,"title":"在不同的姿态和光照条件下的人脸检测","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"OpenCV（Haar-ascade）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我们从OpenCV的Haar-cascade实现开始，它是一个用C语言编写的开源图像处理库。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"优点："},{"type":"text","text":"由于这个库是用C语言编写的，所以它在实时系统中的推理速度非常快。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"缺点："},{"type":"text","text":"这个实现的问题是它无法检测侧脸，而且在不同姿态和光照条件下表现欠佳。"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"MTCNN"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这种算法基于深度学习方法。它使用深度级联卷积神经网络（Deep Cascaded Convolutional Neural Networks）来检测人脸。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"优点："},{"type":"text","text":"它比OpenCV的Haar-Cascade方法准确性更高"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"缺点："},{"type":"text","text":"运行时间较长。"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"YOLOV3"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"YOLO（“You only  look  once”）是用于物体检测的最先进的深度学习算法。它由许多卷积神经网络组成，形成一个深度CNN模型(深度意味着模型架构复杂性很高）。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原始的YOLO模型可以检测80个不同的物体类别，而且检测精度很高，而我们只需要用该模型检测一个物体——人脸。我们在WiderFace（包含"},{"type":"text","marks":[{"type":"strong"}],"text":"393,703个面部标签"},{"type":"text","text":"的图像数据集）数据集上训练了这个算法。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"YOLO算法还有一个微型版本，即Yolo-Tiny。Yolo-Tiny需要的计算时间比较少，但却牺牲了一些准确性。我们用相同的数据集训练了一个Yolo-Tiny模型，其边界框（boundary box）结果并不一致。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"优点："},{"type":"text","text":"非常准确，没有任何缺陷。比MTCNN更快。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"缺点："},{"type":"text","text":"由于具有巨大的深度神经网络层，它需要的计算资源更多。因此，该算法在CPU或移动设备上运行地很慢。在GPU上，它的大型架构也需要耗费更多的VRAM。"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"SSD"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"SSD（Single Shot Detector）也是一个类似YOLO的深度卷积神经网络模型。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"优点："},{"type":"text","text":"良好的准确性。它可以检测各种姿势、光照和遮挡。良好的推理速度。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"缺点："},{"type":"text","text":"比YOLO模型差。虽然推理速度较好，但仍不能满足在CPU、低端GPU或移动设备上运行的要求。"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"BlazeFace"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"就像它的名字一样，它是由谷歌发布的速度极快的人脸检测算法。它接受128x128维的图像输入，推理时间是亚毫秒级，已优化到可以在手机中使用。它速度这么快的原因是："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"YOLO和SSD用来检测大量的类别，而BlazeFace不同，是一个专门的人脸检测器模型。因此BlazeFace的深度卷积神经网络架构比YOLO和SSD的架构小。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"它采用的是深度可分离卷积层（Depthwise Separable Convolution），而不是标准的卷积层，这样就降低了计算量。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"优点："},{"type":"text","text":"非常好的推理速度，且人脸检测的准确率高。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"缺点："},{"type":"text","text":"这个模型的优化目标是对手机摄像头获取的图像进行人脸检测，因此它预期人脸会覆盖图像中的大部分区域，而当人脸尺寸较小时，它的识别效果就是很好。所以，当对闭路电视摄像机获取的（CCTV ，Closed Circuit Tele Vision）图像进行人脸检测时，它表现得并不理想。"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"Faceboxes"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Faceboxes是我们使用的最新的人脸检测算法。与BlazeFace类似，它是一个小型的深度卷积神经网络，只为检测一种类别——人脸而设计。它的推理时间可满足CPU上的实时检测需求。它的准确度可以与Yolo人脸检测算法相媲美，而且，不管图像中的人脸较大还是较小，它都可以精确地检测。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"优点："},{"type":"text","text":"推理速度快，准确性好。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"缺点："},{"type":"text","text":"评估仍在进行中。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"2.特征提取"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在检测到图像中的人脸后，我们对人脸进行裁剪，并将其送入特征提取算法，该算法创建面部嵌入（face-embeddings）——一个代表人脸特征的多维（主要是128或512维）向量。我们使用FaceNet算法来创建面部嵌入。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"嵌入向量代表一个人的面部特征。因此，同一个人的两个不同图像的嵌入向量之间的距离比较接近，而不同人的嵌入向量之间的距离比较远。其中，两个向量之间的距离采用的是欧氏距离。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"3. 面部分类"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在得到面部嵌入向量后，我们训练了一种"},{"type":"link","attrs":{"href":"https:\/\/www.engati.com\/glossary\/classification-algorithm","title":null,"type":null},"content":[{"type":"text","text":"分类算法"}]},{"type":"text","text":"，即K-近邻（K-nearest neighbor，KNN）算法，根据一个人的嵌入向量对其进行分类。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"假设在一个组织中，有1000名员工。我们创建了所有员工的面部嵌入，并使用嵌入向量训练分类算法。该算法以面部嵌入向量作为输入，以人的名字作为输出返回。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在把图片放到网上前，用户可以采用过滤器修改图片中的特定像素。人眼无法察觉这些变化，但它会让面部识别算法觉得很困惑。 —— "},{"type":"link","attrs":{"href":"https:\/\/www.thalesgroup.com\/en\/markets\/digital-identity-and-security\/government\/biometrics\/facial-recognition","title":null,"type":null},"content":[{"type":"text","text":"ThalesGroup"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"当前，面部识别算法已经取得了巨大的进步。但这仅仅是技术革命的开始。可以想象一下，未来面部识别算法和聊天机器人技术的联合起来是多么强大。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"英文原文："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"link","attrs":{"href":"https:\/\/www.pimonk.com\/post\/how-do-facial-recognition-systems-algorithms-work-in-2021","title":null,"type":null},"content":[{"type":"text","text":"https:\/\/www.pimonk.com\/post\/how-do-facial-recognition-systems-algorithms-work-in-2021"}]}]}]}