深度学习图像预处理-python缩放裁剪

使用深度学习进行图像类任务时，通常网络的输入大小是固定的，最近在进行涉及到文字检测的工作中，由于预处理resize缩小了原图，导致字体变模糊，从而检测失败，后来想到使用overlap来对图像进行缩放裁剪，即先将原图缩放到一定尺寸，再裁剪得到网络的输入。

好了，来说正题，使用yolov3，网络的输入是352x352x3，而输入图像大小为几百上千不等，因此需对原图进行resize，起初直接进行缩放 + 填充，检测的map很低，后来分析发现有些352x352的输入图像中的文字已经很模糊，因此直接缩放的方案不可行，改进后方案如下：

1. 原图最大尺寸大于1000，则resize到800x800，再裁剪为9个352x352，overlap为128个像素
2. 原图最大尺寸大小500且小于1000，则resize到600x600，再裁剪为4个352x352，overlap为96个像素
3. 原图最大尺寸小于500，则resize到352x352。

python实现代码如下，使用了PIL、opencv库，将整个目录下的图像全部做缩放裁剪处理，代码包含如下功能：

1. 遍历某一目录的文件
2. opencv进行图像载入及保存
3. opencv进行缩放裁剪
4. PIL进行图像显示

import numpy as np
from PIL import Image
import cv2
import os

#输入bgr通道，并显示图像
def img_show(img):
    img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)      #生成numpy数组
    print(type(img_rgb), img_rgb.shape)
    pil_img = Image.fromarray(img_rgb)
    pil_img.show()

#完成图像的等比缩放及黑色填充
def img_resize(cvImageSrc, net_size, width, height):
    if (width != net_size or height != net_size):
        #宽高的缩放比非常接近时直接进行缩放
        if ((net_size / height - net_size / height) < 0.001):
            det_mat = cv2.resize(cvImageSrc, (net_size, net_size))
            return det_mat
        else:
            new_w = width
            new_h = height
        if (net_size / width < net_size / height):
            new_w = net_size
            new_h = max(1, (height * net_size) / width)
        else:
            new_h = net_size
            new_w = max(1, (width * net_size) / height)
        det_mat = np.zeros((net_size, net_size, 3), dtype="uint8")
        if (new_w == width and new_h == height):
            cvImageSrc.copyTo(det_mat)
        else:
            net_w = int(new_w + 0.5)
            net_h = int(new_h + 0.5)
            if (net_w % 2 == 1):
                net_w = net_w - 1
            if (net_h % 2 == 1):
                net_h = net_h - 1
            #print(net_w, net_h)
            det_matROI = cv2.resize(cvImageSrc, (net_w, net_h))

            base_w = int((net_size - new_w) / 2 + 0.5)
            base_h = int((net_size - new_h) / 2 + 0.5)
            #print(base_h, base_w)
            for c in range(3):
                for j in range(net_h):
                    for i in range(net_w):
                        #print(c, j, i)
                        det_mat[j + base_h - 1, i + base_w - 1, :] = det_matROI[j - 1, i - 1, :]
    else:
        det_mat = cvImageSrc
    return det_mat

baseRoot = "/Users/lemonhe/Documents/CNN/dataset/01-data/"
rootdir = baseRoot + "dataset_test"
list = os.listdir(rootdir)  #列出文件夹下所有的目录与文件
print(len(list))
count = 0

threshold1 = 1000
threshold2 = 500

for i in range(0,len(list)):
    path = os.path.join(rootdir, list[i])
    print(path)
    if os.path.isfile(path):
        img = cv2.imread(path)
        if(img is None):
            print("this is nonetype")
        else:
            height, width, channel = img.shape      #获取图像信息
            print(height, width, channel)
            max_dim = max(height, width)
            #img_show(img)
            if(max_dim > threshold1):
                det_mat = img_resize(img, 800, width, height)
                #img_show(det_mat)
                img11 = np.uint8(det_mat[0:352, 0:352, :])
                img12 = np.uint8(det_mat[0:352, 223:575, :])
                img13 = np.uint8(det_mat[0:352, 447:799, :])
                img21 = np.uint8(det_mat[223:575, 0:352, :])
                img22 = np.uint8(det_mat[223:575, 223:575, :])
                img23 = np.uint8(det_mat[223:575, 447:799, :])
                img31 = np.uint8(det_mat[447:799, 0:352, :])
                img32 = np.uint8(det_mat[447:799, 223:575, :])
                img33 = np.uint8(det_mat[447:799, 447:799, :])
                #print(img13.shape)
                path11 = baseRoot + "test1/img" + str(count) + "_11.jpg"
                path12 = baseRoot + "test1/img" + str(count) + "_12.jpg"
                path13 = baseRoot + "test1/img" + str(count) + "_13.jpg"
                path21 = baseRoot + "test1/img" + str(count) + "_21.jpg"
                path22 = baseRoot + "test1/img" + str(count) + "_22.jpg"
                path23 = baseRoot + "test1/img" + str(count) + "_23.jpg"
                path31 = baseRoot + "test1/img" + str(count) + "_31.jpg"
                path32 = baseRoot + "test1/img" + str(count) + "_32.jpg"
                path33 = baseRoot + "test1/img" + str(count) + "_33.jpg"
                cv2.imwrite(path11, img11)
                cv2.imwrite(path12, img12)
                cv2.imwrite(path13, img13)
                cv2.imwrite(path21, img21)
                cv2.imwrite(path22, img22)
                cv2.imwrite(path23, img23)
                cv2.imwrite(path31, img31)
                cv2.imwrite(path32, img32)
                cv2.imwrite(path33, img33)
            elif(max_dim > threshold2):
                det_mat = img_resize(img, 608, width, height)
                img11 = np.uint8(det_mat[0:352, 0:352, :])
                img12 = np.uint8(det_mat[0:352, 255:607, :])
                img21 = np.uint8(det_mat[255:607, 0:352, :])
                img22 = np.uint8(det_mat[255:607, 255:607, :])
                #img_show(img11)
                #img_show(img12)
                #img_show(img21)net_size
                #img_show(img22)
                path11 = baseRoot + "test1/img" + str(count) + "_11.jpg"
                path12 = baseRoot + "test1/img" + str(count) + "_12.jpg"
                path21 = baseRoot + "test1/img" + str(count) + "_21.jpg"
                path22 = baseRoot + "test1/img" + str(count) + "_22.jpg"
                cv2.imwrite(path11, img11)
                cv2.imwrite(path12, img12)
                cv2.imwrite(path21, img21)
                cv2.imwrite(path22, img22)
            else:
                det_mat = img_resize(img, 352, width, height)
                img_show(det_mat)
                path_352 = baseRoot + "test1/img" + str(count) + ".jpg"
                cv2.imwrite(path_352, np.uint8(det_mat))

    count = count + 1
    print(count)

执行python脚本，结果如下：

输入图片如下，分辨率为719x1280，裁剪后输出9幅352x352的子图，这样就完成了图像的预处理。