diff --git a/demo/mask_detection/python/README.md b/demo/mask_detection/python/README.md index 115be73d294a068f8734a4ba62da59d73524d2a8..a2dc51f4a77468aa63e2fd0d18410def69d6f46f 100644 --- a/demo/mask_detection/python/README.md +++ b/demo/mask_detection/python/README.md @@ -1 +1,62 @@ # 口罩佩戴检测模型Python高性能部署方案 + +百度通过 PaddleHub 开源了业界首个口罩人脸检测及人类模型，该模型可以有效检测在密集人类区域中携带和未携带口罩的所有人脸，同时判断出是否有佩戴口罩。开发者可以通过 PaddleHub 快速体验模型效果、搭建在线服务。 + +本文档主要介绍如何完成基于`python`的口罩佩戴检测预测。 + +主要包含两个步骤： +- [1. PaddleHub导出预测模型](#1-paddlehub导出预测模型) +- [2. 基于python的预测](#2-预测部署编译) + +## 1. PaddleHub导出预测模型 + +#### 1.1 安装 `PaddlePaddle` 和 `PaddleHub` + - `PaddlePaddle`的安装: + 请点击[官方安装文档](https://paddlepaddle.org.cn/install/quick) 选择适合的方式 + - `PaddleHub`的安装: `pip install paddlehub` + - `opencv`的安装: `pip install opencv-python` + - +#### 1.2 从`PaddleHub`导出预测模型 +``` +git clone https://github.com/PaddlePaddle/PaddleHub.git +cd PaddleHub/demo/mask_detection/python/ +python export_model.py +``` +在有网络访问条件下，执行`python export_model.py`导出两个可用于推理部署的口罩模型 +其中`pyramidbox_lite_mobile_mask`为移动版模型, 模型更小，计算量低； +`pyramidbox_lite_server_mask`为服务器版模型，在此推荐该版本模型，精度相对移动版本更高。 +成功执行代码后导出的模型路径结构： +``` +pyramidbox_lite_mobile_mask +| +├── mask_detector # 口罩人脸分类模型 +| ├── __model__ # 模型文件 +│ └── __params__ # 参数文件 +| +└── pyramidbox_lite # 口罩人脸检测模型 + ├── __model__ # 模型文件 + └── __params__ # 参数文件 +``` + +## 2. 基于python的预测 + +### 2.1 执行预测程序 +在终端输入以下命令进行预测: +```bash +python infer.py --models_dir=/path/to/models --img_paths=/path/to/images --video_path=/path/to/video --video_size=size/of/video --use_camera=(False/True) +--use_gpu=(False/True) +``` +参数说明如下: + +| 参数 | 是否必须|含义 | +|-------|-------|----------| +| models_dir | Yes|两个模型路径./pyramidbox_lite_mobile_mask | +| img_paths |img_paths/video_path 二选一|需要预测的图片目录 | +| video_path |img_paths/video_path 二选一|需要预测的视频目录| +| video_size |No|预测视频分辨率大小(w,h) | +| use_camera |No|是否打开摄像头进行预测 | +| use_gpu |No|是否GPU，默认为False| + +##3. 可视化 + + diff --git a/demo/mask_detection/python/export_model.py b/demo/mask_detection/python/export_model.py new file mode 100644 index 0000000000000000000000000000000000000000..7697c4a95c9f5db29c8c8757a1f5364011e8f6f4 --- /dev/null +++ b/demo/mask_detection/python/export_model.py @@ -0,0 +1,27 @@ +# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License" +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import paddlehub as hub + +# Load mask detector module from PaddleHub +module = hub.Module(name="pyramidbox_lite_server_mask") +# Export inference model for deployment +module.processor.save_inference_model("./pyramidbox_lite_server_mask") +print("pyramidbox_lite_server_mask module export done!") + +# Load mask detector (mobile version) module from PaddleHub +module = hub.Module(name="pyramidbox_lite_mobile_mask") +# Export inference model for deployment +module.processor.save_inference_model("./pyramidbox_lite_mobile_mask") +print("pyramidbox_lite_mobile_mask module export done!") diff --git a/demo/mask_detection/python/infer.py b/demo/mask_detection/python/infer.py new file mode 100644 index 0000000000000000000000000000000000000000..be2e96b9ed99d4c988ec1f5957c59e4f445c154f --- /dev/null +++ b/demo/mask_detection/python/infer.py @@ -0,0 +1,274 @@ +# coding: utf8 +# copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +import sys +import ast +import time +import json +import argparse + +import numpy as np +import cv2 + +import paddle.fluid as fluid + +from PIL import Image +from PIL import ImageDraw + +import argparse + + +def parse_args(): + parser = argparse.ArgumentParser('mask detection.') + parser.add_argument( + '--models_dir', type=str, default='', help='path of models.') + parser.add_argument( + '--img_paths', type=str, default='', help='path of images') + parser.add_argument( + '--video_path', type=str, default='', help='path of video.') + parser.add_argument( + '--video_size', type=tuple, default=(1920, 1080), help='size of video.') + parser.add_argument( + '--use_camera', + type=bool, + default=False, + help='switch detect video or camera, default:video.') + parser.add_argument( + '--use_gpu', + type=bool, + default=False, + help='switch cpu/gpu, default:cpu.') + args = parser.parse_args() + return args + + +class FaceResult: + def __init__(self, rect_data, rect_info): + self.rect_info = rect_info + self.rect_data = rect_data + self.class_id = -1 + self.score = 0.0 + + +def VisualizeResult(im, faces): + LABELS = ['NO_MASK', 'MASK'] + COLORS = [(0, 0, 255), (0, 255, 0)] + for face in faces: + label = LABELS[face.class_id] + color = COLORS[face.class_id] + left, right, top, bottom = [int(item) for item in face.rect_info] + label_position = (left, top) + cv2.putText(im, label, label_position, cv2.FONT_HERSHEY_SIMPLEX, 1, + color, 2, cv2.LINE_AA) + cv2.rectangle(im, (left, top), (right, bottom), color, 3) + return im + + +def LoadModel(model_dir, use_gpu=False): + config = fluid.core.AnalysisConfig(model_dir + '/__model__', + model_dir + '/__params__') + if use_gpu: + config.enable_use_gpu(100, 0) + config.switch_ir_optim(True) + else: + config.disable_gpu() + config.disable_glog_info() + config.switch_specify_input_names(True) + config.enable_memory_optim() + return fluid.core.create_paddle_predictor(config) + + +class MaskClassifier: + def __init__(self, model_dir, mean, scale, use_gpu=False): + self.mean = np.array(mean).reshape((3, 1, 1)) + self.scale = np.array(scale).reshape((3, 1, 1)) + self.predictor = LoadModel(model_dir, use_gpu) + self.EVAL_SIZE = (128, 128) + + def Preprocess(self, faces): + h, w = self.EVAL_SIZE[1], self.EVAL_SIZE[0] + inputs = [] + for face in faces: + im = cv2.resize( + face.rect_data, (128, 128), + fx=0, + fy=0, + interpolation=cv2.INTER_CUBIC) + # HWC -> CHW + im = im.swapaxes(1, 2) + im = im.swapaxes(0, 1) + # Convert to float + im = im[:, :, :].astype('float32') / 256.0 + # im = (im - mean) * scale + im = im - self.mean + im = im * self.scale + im = im[np.newaxis, :, :, :] + inputs.append(im) + return inputs + + def Postprocess(self, output_data, faces): + argmx = np.argmax(output_data, axis=1) + for idx in range(len(faces)): + faces[idx].class_id = argmx[idx] + faces[idx].score = output_data[idx][argmx[idx]] + return faces + + def Predict(self, faces): + inputs = self.Preprocess(faces) + if len(inputs) != 0: + input_data = np.concatenate(inputs) + im_tensor = fluid.core.PaddleTensor( + input_data.copy().astype('float32')) + output_data = self.predictor.run([im_tensor])[1] + output_data = output_data.as_ndarray() + self.Postprocess(output_data, faces) + + +class FaceDetector: + def __init__(self, model_dir, mean, scale, use_gpu=False, threshold=0.7): + self.mean = np.array(mean).reshape((3, 1, 1)) + self.scale = np.array(scale).reshape((3, 1, 1)) + self.threshold = threshold + self.predictor = LoadModel(model_dir, use_gpu) + + def Preprocess(self, image, shrink): + h, w = int(image.shape[1] * shrink), int(image.shape[0] * shrink) + im = cv2.resize( + image, (h, w), fx=0, fy=0, interpolation=cv2.INTER_CUBIC) + # HWC -> CHW + im = im.swapaxes(1, 2) + im = im.swapaxes(0, 1) + # Convert to float + im = im[:, :, :].astype('float32') + # im = (im - mean) * scale + im = im - self.mean + im = im * self.scale + im = im[np.newaxis, :, :, :] + return im + + def Postprocess(self, output_data, ori_im, shrink): + det_out = [] + h, w = ori_im.shape[0], ori_im.shape[1] + for out in output_data: + class_id = int(out[0]) + score = out[1] + xmin = (out[2] * w) + ymin = (out[3] * h) + xmax = (out[4] * w) + ymax = (out[5] * h) + wd = xmax - xmin + hd = ymax - ymin + valid = (xmax >= xmin and xmin > 0 and ymax >= ymin and ymin > 0) + if score > self.threshold and valid: + roi_rect = ori_im[int(ymin):int(ymax), int(xmin):int(xmax)] + det_out.append(FaceResult(roi_rect, [xmin, xmax, ymin, ymax])) + return det_out + + def Predict(self, image, shrink): + ori_im = image.copy() + im = self.Preprocess(image, shrink) + im_tensor = fluid.core.PaddleTensor(im.copy().astype('float32')) + output_data = self.predictor.run([im_tensor])[0] + output_data = output_data.as_ndarray() + return self.Postprocess(output_data, ori_im, shrink) + + +def predict_images(args): + detector = FaceDetector( + model_dir=args.models_dir + '/pyramidbox_lite/', + mean=[104.0, 177.0, 123.0], + scale=[0.007843, 0.007843, 0.007843], + use_gpu=args.use_gpu, + threshold=0.7) + + classifier = MaskClassifier( + model_dir=args.models_dir + '/mask_detector/', + mean=[0.5, 0.5, 0.5], + scale=[1.0, 1.0, 1.0], + use_gpu=args.use_gpu) + names = [] + image_paths = [] + for name in os.listdir(args.img_paths): + if name.split('.')[-1] in ['jpg', 'png', 'jpeg']: + names.append(name) + image_paths.append(os.path.join(args.img_paths, name)) + images = [cv2.imread(path, cv2.IMREAD_COLOR) for path in image_paths] + + path = './result' + isExists = os.path.exists(path) + if not isExists: + os.makedirs(path) + for idx in range(len(images)): + im = images[idx] + det_out = detector.Predict(im, shrink=0.7) + classifier.Predict(det_out) + img = VisualizeResult(im, det_out) + cv2.imwrite(os.path.join(path, names[idx] + '.result.jpg'), img) + + +def predict_video(args, im_shape=(1920, 1080), use_camera=False): + if args.use_camera: + capture = cv2.VideoCapture(0) + else: + capture = cv2.VideoCapture(args.video_path) + + detector = FaceDetector( + model_dir=args.models_dir + '/pyramidbox_lite/', + mean=[104.0, 177.0, 123.0], + scale=[0.007843, 0.007843, 0.007843], + use_gpu=args.use_gpu, + threshold=0.7) + + classifier = MaskClassifier( + model_dir=args.models_dir + '/mask_detector/', + mean=[0.5, 0.5, 0.5], + scale=[1.0, 1.0, 1.0], + use_gpu=args.use_gpu) + + path = './result' + isExists = os.path.exists(path) + if not isExists: + os.makedirs(path) + fps = 30 + fourcc = cv2.VideoWriter_fourcc(*'mp4v') + writer = cv2.VideoWriter( + os.path.join(path, 'result.mp4'), fourcc, fps, args.video_size) + import time + start_time = time.time() + index = 0 + while (1): + ret, frame = capture.read() + if not ret: + break + print('detect frame:%d' % (index)) + index += 1 + det_out = detector.Predict(frame, shrink=0.5) + classifier.Predict(det_out) + end_pre = time.time() + im = VisualizeResult(frame, det_out) + writer.write(im) + end_time = time.time() + print("include read time:", (end_time - start_time) / index) + writer.release() + + +if __name__ == "__main__": + args = parse_args() + print(args.models_dir) + if args.img_paths != '': + predict_images(args) + elif args.video_path != '': + predict_video(args)