ax_inference.py

import argparse
import sys
import cv2  
import numpy as np
import time

def preprocess_image(im: np.ndarray, model_input_size: list) -> np.ndarray:
    if len(im.shape) < 3:
        im = im[:, :, np.newaxis]
    im_np = np.transpose(im, (2, 0, 1)).astype(np.float32)
    im_np = np.expand_dims(im_np, axis=0)
    
    _, C, H_ori, W_ori = im_np.shape
    H_target, W_target = model_input_size
    
    x_target = np.linspace(0, W_ori - 1, W_target)
    y_target = np.linspace(0, H_ori - 1, H_target)
    xx_target, yy_target = np.meshgrid(x_target, y_target)
    
    x0 = np.floor(xx_target).astype(np.int32)
    x1 = np.minimum(x0 + 1, W_ori - 1)
    y0 = np.floor(yy_target).astype(np.int32)
    y1 = np.minimum(y0 + 1, H_ori - 1)
    
    wx0 = xx_target - x0
    wx1 = 1 - wx0
    wy0 = yy_target - y0
    wy1 = 1 - wy0
    
    im_interp = np.zeros((1, C, H_target, W_target), dtype=np.float32)
    for c in range(C):
        channel_data = im_np[0, c, :, :]
        top = wx1 * channel_data[y0, x0] + wx0 * channel_data[y0, x1]
        bottom = wx1 * channel_data[y1, x0] + wx0 * channel_data[y1, x1]
        im_interp[0, c, :, :] = wy1 * top + wy0 * bottom
    
    image = (im_interp / 1.0).astype(np.uint8)
    
    return image

def postprocess_image(result: np.ndarray, im_size: list)-> np.ndarray:
    result_np = np.squeeze(result, axis=0)
    C, H_ori, W_ori = result_np.shape
    H_target, W_target = im_size  # 目标尺寸(H,W)

    x_target = np.linspace(0, W_ori - 1, W_target)
    y_target = np.linspace(0, H_ori - 1, H_target)
    xx_target, yy_target = np.meshgrid(x_target, y_target)

    x0 = np.floor(xx_target).astype(np.int32)
    x1 = np.minimum(x0 + 1, W_ori - 1)
    y0 = np.floor(yy_target).astype(np.int32)
    y1 = np.minimum(y0 + 1, H_ori - 1)

    wx0 = xx_target - x0
    wx1 = 1 - wx0
    wy0 = yy_target - y0
    wy1 = 1 - wy0

    result_interp = np.zeros((C, H_target, W_target), dtype=np.float32)
    for c in range(C):
        channel_data = result_np[c, :, :]
        top = wx1 * channel_data[y0, x0] + wx0 * channel_data[y0, x1]
        bottom = wx1 * channel_data[y1, x0] + wx0 * channel_data[y1, x1]
        result_interp[c, :, :] = wy1 * top + wy0 * bottom
    
    ma = np.max(result_interp)
    mi = np.min(result_interp)

    result_norm = (result_interp - mi) / (ma - mi + 1e-8)  # 加极小值避免除零
    result_scaled = result_norm * 255
    im_array = np.transpose(result_scaled, (1, 2, 0)).astype(np.uint8)
    im_array = np.squeeze(im_array)
    return im_array

def inference(img_path,
              model_path,
              save_path):
    
    if model_path.endswith(".axmodel"):
        import axengine as ort

    session = ort.InferenceSession(model_path)
    input_name = None
    for inp_meta in session.get_inputs():
        input_shape = inp_meta.shape[2:]
        input_name = inp_meta.name
        print(f"输入名称：{input_name}，输入尺寸：{input_shape}")
    
    model_input_size = [1024, 1024]
    orig_im_bgr = cv2.imread(img_path)
    if orig_im_bgr is None:
        raise FileNotFoundError(f"无法读取图片文件：{img_path}，请检查路径是否正确或图片是否损坏")
    orig_im = cv2.cvtColor(orig_im_bgr, cv2.COLOR_BGR2RGB)  # 转换为RGB格式 (H,W,3)
    orig_im_size = orig_im.shape[0:2]  
    image = preprocess_image(orig_im, model_input_size)
    
    t1 = time.time()
    result = session.run(None, {input_name: image})
    t2 = time.time()
    print(f"推理时间：{(t2-t1)*1000:.2f} ms")
   
    result_image = postprocess_image(result[0], orig_im_size)  # 得到单通道掩码 (H,W)
    orig_im_unchanged = cv2.imread(img_path, cv2.IMREAD_UNCHANGED)  # 读取所有通道（BGR/BGRA）
    mask = result_image  # 单通道掩码 (H,W)，值范围0-255
    if orig_im_unchanged.shape[-1] == 3:  # 原图为BGR格式（无透明通道）
        b, g, r = cv2.split(orig_im_unchanged)
        a = mask  
        no_bg_image = cv2.merge((b, g, r, a))  # 合并为BGRA格式
    elif orig_im_unchanged.shape[-1] == 4:  # 原图为BGRA格式（已有透明通道）
        b, g, r, _ = cv2.split(orig_im_unchanged)
        a = mask
        no_bg_image = cv2.merge((b, g, r, a))
    else:
        raise ValueError(f"不支持的图片通道数：{orig_im_unchanged.shape[-1]}，仅支持3通道（BGR）或4通道（BGRA）")
    
    if save_path.lower().endswith(('.jpg', '.jpeg')):
        cv2.imwrite(save_path, cv2.cvtColor(no_bg_image, cv2.COLOR_BGRA2BGR))
        print(f"JPG格式不支持透明通道，已丢弃Alpha通道，结果保存至：{save_path}")
    else:
        cv2.imwrite(save_path, no_bg_image)
        print(f"推理完成，结果已保存至：{save_path}")

def parse_args() -> argparse.ArgumentParser:
    parser = argparse.ArgumentParser(description="ax rmbg exsample")
    parser.add_argument("--model","-m", type=str, help="compiled.axmodel path")
    parser.add_argument("--img","-i", type=str, help="img path")
    parser.add_argument("--save_path", type=str, default="./result.png", help="save result path (png)")

    args = parser.parse_args()
    return args

if __name__ == "__main__":
    args = parse_args()
    
    print(f"Command: {' '.join(sys.argv)}")
    print("Parameters:")
    print(f"  --model: {args.model}")
    print(f"  --img_path: {args.img}")
    print(f"  --save_path: {args.save_path}")

    inference(args.img, args.model, args.save_path)