一种大图片裁剪方法

在裁剪大图片的时候，存在不能完全切割的问题，将大图片切成小图片制作训练集时，不能完全切割会造成对图像信息的浪费。我简单写了一个可以将图片完全切割的代码。

示例：

原图大小：3000×4000×3（height × width × channels）

目标大小：512×512×3 （height × width × channels）

图-1 原图

 代码：

#Part1：导入相关模块
import cv2
from tqdm import tqdm
import os

#Part2：设置裁剪目标图片大小，设置输入输出位置
size = 512
out_image_path = "D:/out_image_path/"
out_label_path = "D:/out_label_path/"
in_image_path = "D:/in_image_path/"
in_label_path = "D:/in_label_path/"

#Part3：设置GPU
os.environ["CUDA_VISIBLE_DEVICES"] = "0"

#Part4：读取图片和标签数量
origin_images_path = []
for i in os.listdir(in_image_path):
    origin_images_path.append(in_image_path + i)
origin_labels_path = []
for j in os.listdir(in_label_path):
    origin_labels_path.append(in_label_path + j)

#Part4：裁剪函数
def generate_train_dataset(out_image_path, out_label_path, input_height, input_width):
    p_count = 0 #裁剪后图片数量统计
    images_path = origin_images_path
    labels_path = origin_labels_path
    main_height_num = input_height // size #以整除的形式算出能够完整不重复切出目标大小的区域
    main_width_num = input_width // size #以整除的形式算出能够完整不重复切出目标大小的区域
    main_count = main_width_num*main_height_num #将不重复完整切割区域定义为主切割区

    for i in tqdm(range(len(images_path))):
        count_m = 0 #主切割区切图数量统计
        count_h = 0 #下边缘边角区切图数量统计
        count_w = 0 #右边缘边角区切图数量统计
        image = cv2.imread(images_path[i], 1) #根据需要设置参数值，如变化，注意下方切图通道数
        label = cv2.imread(labels_path[i], 1) #根据需要设置参数值，如变化，注意下方切图通道数
        X_height, X_width = image.shape[0], image.shape[1]
        width_start = 0
        height_start = 0
        while count_m < main_count:
            image_ogi = image[height_start: height_start + size, width_start: width_start + size, :]
            label_ogi = label[height_start: height_start + size, width_start: width_start + size, :]
            cv2.imwrite((out_image_path + '%d_img.jpg' % p_count), image_ogi) #注意目标图片格式
            cv2.imwrite((out_label_path + '%d_lab.png' % p_count), label_ogi) #注意目标图片格式
            count_m += 1
            p_count += 1
            width_start += size
            if width_start == size * main_width_num :
                height_start = height_start +size
                width_start = 0

        width_start = 0
        height_start = X_height - size

        while count_w < main_width_num + 1 :
            if X_width - width_start < size:
                width_start = X_width - size
            image_ogi = image[height_start: height_start + size, width_start: width_start + size, :]
            label_ogi = label[height_start: height_start + size, width_start: width_start + size, :]
            cv2.imwrite((out_image_path + '%d_img.jpg' % p_count), image_ogi)
            cv2.imwrite((out_label_path + '%d_lab.png' % p_count), label_ogi)
            count_w += 1
            p_count += 1
            width_start += size

        width_start = X_width - size
        height_start = 0

        while count_h < main_height_num: #注意区别上一个while循环，这里的循环数比上一个少1，避免最右下角被重复切图
            if X_height - height_start < size:
                height_start = X_height - size
            image_ogi = image[height_start: height_start + size, width_start: width_start + size, :]
            label_ogi = label[height_start: height_start + size, width_start: width_start + size, :]
            cv2.imwrite((out_image_path + '%d_img.jpg' % p_count), image_ogi)
            cv2.imwrite((out_label_path + '%d_lab.png' % p_count), label_ogi)
            count_h += 1
            p_count += 1
            height_start += size

#调用
if __name__ == '__main__':
    if not os.path.exists(out_image_path): os.mkdir(out_image_path)
    if not os.path.exists(out_label_path): os.mkdir(out_label_path)
    generate_train_dataset(out_image_path=out_image_path, out_label_path=out_label_path,input_height=3000,input_width=4000)
    #设置参数

 图-2 裁剪后

自动计算生成48张裁剪后图片。