5.AI应用样例体验

       在Orangepi AIPRO板卡系统中还内置了多个有趣的AI处理案例，大大降低了学习门槛。

       在人像分割场景中，Orangepi AIPRO可以通过训练好的模型轻松实现深度学习神经网络PortraitNet，（PortraitNet是实时人像分割模型，该模型可以在移动设备上有效且高效地运行，基于轻量级的 U 形架构，在训练阶段有两个辅助损失，而在测试阶段不需要额外的成本进行肖像推理。两个辅助损失是边界损失和一致性约束损失。前者提高了边界像素的精度，后者增强了复杂光照环境下的鲁棒性）。

      5.1. 打开内置Jupyter Notebook AI样例的操作很便捷，只需要下面两条命令；

cd samples/notebooks/
./start_notebook.sh

 5.2. 运行Jupyter Notebook AI样例

    5.3.  输出对比图像

class Seg(object):
    """人像分割模型推理"""
    def __init__(self, model_path, model_width, model_height):
        self._model_path = model_path
        self._model_width = model_width
        self._model_height = model_height
        self.device_id = 0
        self._dvpp = None
        self._model = None

    def init(self):
        """初始化相关资源"""
        # Initialize dvpp
        self._dvpp = AclLiteImageProc()

        # Load model
        self._model = AclLiteModel(self._model_path)

        return const.SUCCESS

    @utils.display_time
    def pre_process(self, image):
        """图片预处理"""
        image_dvpp = image.copy_to_dvpp()
        yuv_image = self._dvpp.jpegd(image_dvpp)
        resized_image = self._dvpp.resize(yuv_image,
                                          self._model_width, self._model_height)
        return resized_image     

    @utils.display_time
    def inference(self, input_data):
        """模型推理"""
        return self._model.execute(input_data)

    @utils.display_time
    def post_process(self, infer_output, image_name):
        """获取分割结果"""
        data = infer_output[0]
        vals = data.flatten()
        mask = np.clip((vals * 255), 0, 255)
        mask = mask.reshape(224, 224, 2)
        cv2.imwrite(os.path.join(MASK_DIR, image_name), mask[:, :, 0])
        return mask 

    
@utils.display_time
def background_replace(bg_path, ori_path, mask_path):
    """将人像分割结果与背景图片结合"""
    background = cv2.imread(bg_path)
    height, width = background.shape[:2]
    ori_img = cv2.imread(ori_path)
    mask = cv2.imread(mask_path, 0)
    mask = mask / 255
    mask_resize = cv2.resize(mask, (width, height))
    ori_img = cv2.resize(ori_img, (width, height))
    mask_bg = np.repeat(mask_resize[..., np.newaxis], 3, 2)
    result = np.uint8(background * mask_bg + ori_img * (1 - mask_bg))
    cv2.imwrite(os.path.join(OUTPUT_DIR, os.path.basename(mask_path)), result)

    
def main():
    """推理主函数"""
    os.makedirs(OUTPUT_DIR, exist_ok=True)
    os.makedirs(MASK_DIR, exist_ok=True)

    acl_resource = AclLiteResource()
    acl_resource.init()

    seg = Seg(MODEL_PATH, MODEL_WIDTH, MODEL_HEIGHT)
    ret = seg.init()
    utils.check_ret("seg.init ", ret)

    images_list = [os.path.join(IMAGE_DIR, img)
                   for img in os.listdir(IMAGE_DIR)
                   if os.path.splitext(img)[1] in const.IMG_EXT]

    for image_file in images_list:
        image_name = os.path.basename(image_file)
        if image_name != 'background.jpg':
            print('====' + image_name + '====')
            # read image
            image = AclLiteImage(image_file)
            # Preprocess the picture
            resized_image = seg.pre_process(image)
            # Inference
            result = seg.inference([resized_image, ])
            # Post-processing
            mask = seg.post_process(result, image_name)
            # Fusion of segmented portrait and background image
            background_replace(os.path.join(IMAGE_DIR, 'background.jpg'), \
                                        image_file, os.path.join(MASK_DIR, image_name))

• Orangepi AIPRO 官方网站
• Orangepi AIPRO Wiki
• Orangepi AIPRO 硬件文档
• OrangePI AIPRO 用户手册