import cv2 import numpy as np import onnxruntime as ort class QualityOfClarity: def __init__(self, low_thresh=0.10, high_thresh=0.20): self.low_thresh = low_thresh self.high_thresh = high_thresh def reblur(self, data, width, height): data = np.array(data, dtype=np.float32).reshape((height, width)) # 创建一维核 kernel = np.ones((9,), np.float32) / 9.0 # 垂直方向模糊处理 BVer = cv2.filter2D(data, -1, kernel.reshape(-1, 1), borderType=cv2.BORDER_REPLICATE) # 水平方向模糊处理 BHor = cv2.filter2D(data, -1, kernel.reshape(1, -1), borderType=cv2.BORDER_REPLICATE) s_FVer, s_FHor, s_Vver, s_VHor = 0.0, 0.0, 0.0, 0.0 # 计算垂直方向的差分 D_Fver = np.abs(data[1:, :] - data[:-1, :]) D_BVer = np.abs(BVer[1:, :] - BVer[:-1, :]) # 计算垂直方向的累积 s_FVer = np.sum(D_Fver) s_Vver = np.sum(np.maximum(0.0, D_Fver - D_BVer)) # 计算水平方向的差分 D_FHor = np.abs(data[:, 1:] - data[:, :-1]) D_BHor = np.abs(BHor[:, 1:] - BHor[:, :-1]) # 计算水平方向的累积 s_FHor = np.sum(D_FHor) s_VHor = np.sum(np.maximum(0.0, D_FHor - D_BHor)) b_FVer = (s_FVer - s_Vver) / s_FVer b_FHor = (s_FHor - s_VHor) / s_FHor blur_val = max(b_FVer, b_FHor) return blur_val def grid_max_reblur(self, img, rows, cols): height, width = img.shape row_height = height // rows col_width = width // cols blur_val = float('-inf') for y in range(rows): for x in range(cols): grid = img[y * row_height: (y + 1) * row_height, x * col_width: (x + 1) * col_width] this_grad_blur_val = self.reblur(grid, col_width, row_height) if this_grad_blur_val > blur_val: blur_val = this_grad_blur_val return max(blur_val, 0.0) def clarity_estimate(self, image): # x, y, w, h = rect # if w < 9 or h < 9: # return 0.0 src_data = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) # src_data = gray_data[y:y+h, x:x+w] blur_val = self.grid_max_reblur(src_data, 2, 2) clarity = 1.0 - blur_val T1, T2 = 0.0, 1.0 if clarity <= T1: clarity = 0.0 elif clarity >= T2: clarity = 1.0 else: clarity = (clarity - T1) / (T2 - T1) return clarity def check(self, image): clarity = self.clarity_estimate(image) if clarity < self.low_thresh: level = "LOW" elif self.low_thresh <= clarity < self.high_thresh: level = "MEDIUM" else: level = "HIGH" return level != 'LOW' # return {'level': level, "score": clarity} class QualityChecker: # check resolution and clarity of image def __init__(self, v0=70.0, v1=100.0, v2=210.0, v3=230.0, hw = (112,112)): self.bright_thresh0 = v0 self.bright_thresh1 = v1 self.bright_thresh2 = v2 self.bright_thresh3 = v3 self.middle_thresh = (self.bright_thresh1 + self.bright_thresh2) / 2 self.rolu_thrds = hw # （h，w） def get_bright_score(self, bright): bright_score = 1.0 / (abs(bright - self.middle_thresh) + 1) return bright_score def grid_max_bright(self, img, rows, cols): row_height = img.shape[0] // rows col_width = img.shape[1] // cols # 使用列表生成式获取所有网格的平均亮度 grid_means = [ np.mean(img[y * row_height:(y + 1) * row_height, x * col_width:(x + 1) * col_width]) for y in range(rows) for x in range(cols) ] # 获取最大亮度值 bright_val = max(grid_means) return max(bright_val, 0) def check_bright(self, face_image): gray = cv2.cvtColor(face_image, cv2.COLOR_BGR2GRAY) if len(face_image.shape) == 3 else face_image bright_value = self.grid_max_bright(gray, 3, 3) if bright_value < self.bright_thresh0 or bright_value > self.bright_thresh3: level = "LOW" elif (self.bright_thresh0 <= bright_value < self.bright_thresh1) or (self.bright_thresh2 < bright_value <= self.bright_thresh3): level = "MEDIUM" else: level = "HIGH" return level == "HIGH" def check_resolution(self, face_image): H, W = face_image.shape[:2] if H < self.rolu_thrds[0] or W < self.rolu_thrds[1]: return False return True class QualityOfPose: def __init__(self, pad=0.3, yaw_thrd=30, pitch_thrd=25, var_onnx_path = './checkpoints/fsanet-var.onnx', conv_onnx_path='./checkpoints/fsanet-conv.onnx') -> None: self.pad = pad self.input_width = 64 self.input_height = 64 self.yaw_thrd = abs(yaw_thrd) self.pitch_thrd = abs(pitch_thrd) self.var_fsanet = ort.InferenceSession(var_onnx_path) self.conv_fsanet = ort.InferenceSession(conv_onnx_path) self.var_input_names = [input_.name for input_ in self.var_fsanet.get_inputs()] self.var_output_names = [output.name for output in self.var_fsanet.get_outputs()] self.conv_input_names = [input_.name for input_ in self.conv_fsanet.get_inputs()] self.conv_output_names = [output.name for output in self.conv_fsanet.get_outputs()] def transform(self, mat): h, w = mat.shape[:2] nh = int(h + self.pad * h) nw = int(w + self.pad * w) nx1 = max(0, (nw - w) // 2) ny1 = max(0, (nh - h) // 2) # Create a padded canvas and copy the image into the center canvas = np.zeros((nh, nw, 3), dtype=np.uint8) canvas[ny1:ny1 + h, nx1:nx1 + w] = mat # Resize the image to the input dimensions canvas = cv2.resize(canvas, (self.input_width, self.input_height)) # Normalize the image in-place canvas = canvas.astype(np.float32) mean = 127.5 scale = 1.0 / 127.5 canvas = (canvas - mean) * scale # Convert to CHW format canvas = np.transpose(canvas, (2, 0, 1)) # Create a tensor input_tensor = np.expand_dims(canvas, axis=0).astype(np.float32) return input_tensor def detect_angle(self, img): input_tensor = self.transform(img) var_output = self.var_fsanet.run( self.var_output_names, {self.var_input_names[0]: input_tensor} )[0] conv_output = self.conv_fsanet.run( self.conv_output_names, {self.conv_input_names[0]: input_tensor} )[0] yaw, pitch, roll = np.mean(np.vstack((var_output,conv_output)), axis=0) return yaw, pitch, roll def check(self, image): yaw, pitch, roll = self.detect_angle(image) if abs(yaw) <= self.yaw_thrd and abs(pitch) <= self.pitch_thrd: return "frontFace" elif yaw < -1.0 * self.yaw_thrd: return "rightFace" elif yaw > self.yaw_thrd: return "leftFace" elif pitch > self.pitch_thrd: return "upFace" elif pitch < -1.0 * self.pitch_thrd: return "downFace" else: return "implementError" if __name__ == "__main__": qp = QualityOfPose() img = cv2.imread("/home/bns/liteAitoolkit/Demos_onnx/examples/det_oeyecmouth.jpg") angles = qp.check(img) print(f"ONNXRuntime Version yaw: {angles[0]} pitch: {angles[1]} roll: {angles[2]}") pass