【深度学习实战（4）】使用PIL库实现自己的letter_box操作和逆letter_box操作

一、letter_box，逆letter_box

深度学习模型输入图片的尺寸为正方形，而数据集中的图片一般为长方形，粗暴的resize会使得图片失真，采用letterbox可以较好的解决这个问题。该方法可以保持图片的长宽比例，剩下的部分采用灰色填充。对于语义分割任务，我们从模型推理出来的mask图尺寸和模型输入尺寸一致，要想获取原始的图像尺寸的预测结果，这时候就还需要逆letter_box操作。

二、代码

本例中，模型输入尺寸为604x640，而我们读取的图片的实际尺寸为128x384，通过letter_box操作，实现将原始图像以不失真的方式调整为640x640。再通过逆letter_box操作，使得尺寸再恢复为128x384。

from PIL import Image
import numpy as np
import cv2

# ------------------------------------------------------------------#
#   letter_box操作
# ------------------------------------------------------------------#
def letter_box_pil(orgin_image, target_size):
    orgin_image.show()
    orgin_w, orgin_h  = orgin_image.size
    target_w, target_h = target_size

    # ------------------------------------------------------------------#
    #   以原图较长的边为准，将长边resize到target_size大小；原图较短的边, 进行等比缩放
    # ------------------------------------------------------------------#
    scale   = min(target_w/orgin_w, target_h/orgin_h)
    new_w = int(orgin_w*scale)
    new_h = int(orgin_h*scale)

    image   = orgin_image.resize((new_w,new_h), Image.BICUBIC)

    # ------------------------------------------------------------------#
    #   构建输出图像
    # ------------------------------------------------------------------#
    target_image = Image.new('RGB', target_size, (128,128,128))
    target_image.paste(image, ((target_w-new_w)//2, (target_h-new_h)//2))
    target_image.show("pil_image_letter_box")
    return target_image, new_w,new_h,orgin_w,orgin_h

# ------------------------------------------------------------------#
#   逆letter_box操作
# ------------------------------------------------------------------#
def letter_box_pil_inv(target_image,target_size,new_h,new_w):
    # ------------------------------------------------------------------#
    #   PIL.Image -> array
    # ------------------------------------------------------------------#
    arr_image = np.array(target_image)

    #--------------------------------------#
    #   将灰条部分截取掉
    #--------------------------------------#
    arr_image = arr_image[int((target_size[0] - new_h) // 2) : int((target_size[0] - new_h) // 2 + new_h), \
        int((target_size[1] - new_w) // 2) : int((target_size[1] - new_w) // 2 + new_w)]
    # ------------------------------------------------------------------#
    #   进行图片的resize
    # ------------------------------------------------------------------#
    arr_image = cv2.resize(arr_image, (orgin_w, orgin_h), interpolation = cv2.INTER_LINEAR)

    # ------------------------------------------------------------------#
    #   array -> PIL.Image
    # ------------------------------------------------------------------#
    pil_image = Image.fromarray(arr_image)
    pil_image.show("pil_image_letter_box_inv")
    return pil_image


orgin_image = Image.open('demo.jpg')
target_size = [640, 640]

# ------------------------------------------------------------------#
#   letter_box操作
# ------------------------------------------------------------------#
target_image, new_w, new_h,orgin_w,orgin_h = letter_box_pil(orgin_image, target_size)

# ------------------------------------------------------------------#
#   逆letter_box操作
# ------------------------------------------------------------------#
image = letter_box_pil_inv(target_image,target_size,new_h,new_w)

运行结果