对mtcnn的人脸对齐的理解

概念理解

人脸识的流程&＃xff1a;人脸检测 ——人脸对齐——特征提取——相似度对比

人脸对齐也是关键的一步&＃xff0c;在不同的应用场景下&＃xff0c;会直接影响到人脸识别的结果。因为是否进行人脸对齐&＃xff0c;会影响到提取到的特征&＃xff0c;对齐前后提取到的特征是有差别的。

人脸对齐&＃xff08;矫正&＃xff09;&＃xff1a;就是检测到人脸角度不正&＃xff0c;关键点不对齐&＃xff0c;然后需要对齐操作。

人脸对齐前后的效果对比如下图&＃xff0c;发现对齐后效果还是挺好的。

那么我们要怎么实施人脸对齐呢&＃xff1f;大致的思路是&＃xff1a;先设定一个src作为标准的人脸关键点的位置&＃xff0c;然后和我们检测到的人脸关键点dst进行相似变换&＃xff0c;变换的过程包括旋转、平移、缩放&＃xff0c;这样就得到一个齐次变换矩阵M&＃xff0c;然后把M作为参数进行仿射变换得到对齐后的人脸图片。

代码

可以结合下代码做更好的理解&＃xff1a;

def preprocess(img, bbox&＃61;None, landmark&＃61;None, **kwargs):if isinstance(img, str): # 判断一个对象是否是一个已知的类型&＃xff0c;类似type()img &＃61; read_image(img, **kwargs)M &＃61; Noneimage_size &＃61; []str_image_size &＃61; kwargs.get(&＃39;image_size&＃39;, &＃39;&＃39;)if len(str_image_size)>0: # 得到图片的image_size&＃xff0c;这里用112x112image_size &＃61; [int(x) for x in str_image_size.split(&＃39;,&＃39;)]if len(image_size)&＃61;&＃61;1:image_size &＃61; [image_size[0], image_size[0]]assert len(image_size)&＃61;&＃61;2assert image_size[0]&＃61;&＃61;112assert image_size[0]&＃61;&＃61;112 or image_size[1]&＃61;&＃61;96if landmark is not None: # 如果landmark不为none&＃xff0c;就计算出Massert len(image_size)&＃61;&＃61;2src &＃61; np.array([ # 人脸的5个关键点的位置&＃xff0c;是固定的[30.2946, 51.6963],[65.5318, 51.5014],[48.0252, 71.7366],[33.5493, 92.3655],[62.7299, 92.2041] ], dtype&＃61;np.float32 )if image_size[1]&＃61;&＃61;112: # 如果为112&＃xff0c;则要把这些坐标的横坐标都加上8.0src[:,0] &＃43;&＃61; 8.0 # 那么8.0是怎么计算的呢&＃xff1f;(112-96)/2 &＃61; 8.0dst &＃61; landmark.astype(np.float32) # 目标关键点&＃xff0c;设置一下它的数据类型tform &＃61; trans.SimilarityTransform() # 引用 class SimilarityTransform()tform.estimate(dst, src) # 从一组对应的点估计转换M &＃61; tform.params[0:2,:] # 得到(3, 3) 的齐次变换矩阵#M &＃61; cv2.estimateRigidTransform( dst.reshape(1,5,2), src.reshape(1,5,2), False)if M is None: # 如果通过上面的变换没有找到齐次变换矩阵&＃xff0c;就用以下的方法来调整bboxif bbox is None: #use center crop # 如果没有bbox&＃xff0c;用中心来进行裁剪det &＃61; np.zeros(4, dtype&＃61;np.int32)det[0] &＃61; int(img.shape[1]*0.0625)det[1] &＃61; int(img.shape[0]*0.0625)det[2] &＃61; img.shape[1] - det[0]det[3] &＃61; img.shape[0] - det[1]else: # 直接使用bboxdet &＃61; bboxmargin &＃61; kwargs.get(&＃39;margin&＃39;, 44) # margin的值一般为0.2&＃xff0c;表示两个类之间的间距bb &＃61; np.zeros(4, dtype&＃61;np.int32) # 4个关键点坐标bb[0] &＃61; np.maximum(det[0]-margin/2, 0)bb[1] &＃61; np.maximum(det[1]-margin/2, 0)bb[2] &＃61; np.minimum(det[2]&＃43;margin/2, img.shape[1])bb[3] &＃61; np.minimum(det[3]&＃43;margin/2, img.shape[0])ret &＃61; img[bb[1]:bb[3],bb[0]:bb[2],:] # 得到4个关键点坐标if len(image_size)>0:ret &＃61; cv2.resize(ret, (image_size[1], image_size[0])) # 图片缩放到112return ret else: #do align using landmarkassert len(image_size)&＃61;&＃61;2#src &＃61; src[0:3,:]#dst &＃61; dst[0:3,:]#print(src.shape, dst.shape)#print(src)#print(dst)#print(M)warped &＃61; cv2.warpAffine(img,M,(image_size[1],image_size[0]), borderValue &＃61; 0.0) # 进行仿射变换#tform3 &＃61; trans.ProjectiveTransform()#tform3.estimate(src, dst)#warped &＃61; trans.warp(img, tform3, output_shape&＃61;_shape)return warped

class SimilarityTransform(EuclideanTransform):"""2D similarity transformation of the form:X &＃61; a0 * x - b0 * y &＃43; a1 &＃61;&＃61; s * x * cos(rotation) - s * y * sin(rotation) &＃43; a1Y &＃61; b0 * x &＃43; a0 * y &＃43; b1 &＃61;&＃61; s * x * sin(rotation) &＃43; s * y * cos(rotation) &＃43; b1where &＃96;&＃96;s&＃96;&＃96; is a scale factor and the homogeneous transformation matrix is::[[a0 b0 a1][b0 a0 b1][0 0 1]]The similarity transformation extends the Euclidean transformation with asingle scaling factor in addition to the rotation and translationparameters.Parameters----------matrix : (3, 3) array, optionalHomogeneous transformation matrix.scale : float, optionalScale factor.rotation : float, optionalRotation angle in counter-clockwise direction as radians.translation : (tx, ty) as array, list or tuple, optionalx, y translation parameters.Attributes----------params : (3, 3) arrayHomogeneous transformation matrix."""def __init__(self, matrix&＃61;None, scale&＃61;None, rotation&＃61;None,translation&＃61;None):params &＃61; any(param is not Nonefor param in (scale, rotation, translation))if params and matrix is not None:raise ValueError("You cannot specify the transformation matrix and"" the implicit parameters at the same time.")elif matrix is not None:if matrix.shape !&＃61; (3, 3):raise ValueError("Invalid shape of transformation matrix.")self.params &＃61; matrixelif params:if scale is None:scale &＃61; 1if rotation is None:rotation &＃61; 0if translation is None:translation &＃61; (0, 0)self.params &＃61; np.array([[math.cos(rotation), - math.sin(rotation), 0],[math.sin(rotation), math.cos(rotation), 0],[ 0, 0, 1]])self.params[0:2, 0:2] *&＃61; scaleself.params[0:2, 2] &＃61; translationelse:# default to an identity transformself.params &＃61; np.eye(3)def estimate(self, src, dst):"""Estimate the transformation from a set of corresponding points.You can determine the over-, well- and under-determined parameterswith the total least-squares method.Number of source and destination coordinates must match.Parameters----------src : (N, 2) arraySource coordinates.dst : (N, 2) arrayDestination coordinates.Returns-------success : boolTrue, if model estimation succeeds."""self.params &＃61; _umeyama(src, dst, True)return True&＃64;propertydef scale(self):if abs(math.cos(self.rotation))

2维相似度变换公式&＃xff1a;

公式中s是缩放因子&＃xff0c;齐次变换矩阵是

[[a0  b0  a1]
 [b0  a0  b1]
 [0   0    1   ]]

参数&＃xff1a;

matrix : (3, 3) 数组&＃xff0c;可选的齐次变换矩阵

scale : 缩放因子

rotation : 逆时针旋转角度为弧度

translation : (tx, ty) 是一个 array, list or tuple, 转换参数

params : (3, 3) 数组&＃xff0c;齐次变换矩阵

除了旋转和平移参数外&＃xff0c;相似变换还扩展了具有单个比例因子的欧几里得变换。从一组相应的点估计转换&＃xff0c;可以使用总最小二乘法确定过、好和欠的参数&＃xff0c;且要求源坐标和目标坐标的数量必须匹配。