python 實(shí)現(xiàn)從高分辨圖像上摳取圖像塊

#coding=utf-8
import cv2
import numpy as np
import os
# 程序?qū)崿F(xiàn)功能：
# 根據(jù)patch在高分辨率圖像上的索引值，crop出對(duì)應(yīng)區(qū)域的圖像
# 并驗(yàn)證程序的正確性
'''
對(duì)于當(dāng)前輸入的3328*3328的高分辨率特征圖，首先resize到640*640
然后根據(jù)當(dāng)前的patch文件名（包含了patch在高分辨率圖像上的行索引和列索引）
這個(gè)索引值是將高分辨率圖像劃分成多個(gè)沒有overlap的256*256的圖像塊之后的行索引和列索引
行索引range(1,11)，列索引range(0,12)
3328=13*256
'''

index='IDRiD_03_3_12.jpg'
raw_img_path='F:\\2\\eye_seg_con\\eye_seg\\joint_data\\raw_image\\train'
patches_path='F:\\2\\eye_seg_con\\eye_seg\\joint_data\\patches\\train'
true_patches=cv2.imread(os.path.join(patches_path,index))[:,:,::-1]

print(os.path.join(raw_img_path,index.split('_')[0]+index.split('_')[1]+'.jpg'))

hr_img=cv2.imread(os.path.join(raw_img_path,index.split('_')[0]+'_'+index.split('_')[1]+'.jpg'))[:,:,::-1]
hr_img=cv2.resize(hr_img,(640,640))# hr_img RGB

'''
640/13=49.23076923076923 記作unit
將640*640的區(qū)域平均劃分成13*13份，每一份的像素點(diǎn)大小是unit*unit
然后將對(duì)應(yīng)位置（取整）的圖像塊摳出來，resize成256*256大小
'''
unit=640/13
patch_row_num = int(index[:-4].split('_')[2])
patch_col_num = int(index[:-4].split('_')[3])

row_start=round(patch_row_num*unit)
row_end=round((patch_row_num+1)*unit)
col_start=round(patch_col_num*unit)
col_end=round((patch_col_num+1)*unit)

my_patch=hr_img[row_start:row_end,col_start:col_end,:]
my_patch=cv2.resize(my_patch,(256,256))
my_patch=np.array(my_patch,dtype=np.uint8)

cv2.imshow('true_patches',true_patches[:,:,::-1])
cv2.waitKey(0)

cv2.imshow('my_patch',my_patch[:,:,::-1])
cv2.waitKey(0)

# # hr_img RGB
#
# # cv2.imshow('1',hr_img[:,:,::-1])
# # cv2.waitKey(0)
#
# hr_img2=cv2.imread(os.path.join(raw_img_path,index.split('_')[0]+'_'+index.split('_')[1]+'.jpg'))
# hr_img2=cv2.resize(hr_img2,(640,640))[:,:,::-1]# hr_img2 RGB
# # cv2.imshow('2',hr_img2[:,:,::-1])
# # cv2.waitKey(0)
#
# print(np.sum(hr_img2-hr_img))# 0

# 結(jié)論：
# 對(duì)于cv2.resize函數(shù)而言,無論是先進(jìn)行BGR的通道轉(zhuǎn)換,再resize,還是先進(jìn)行resize,再進(jìn)行BGR通道轉(zhuǎn)換
# 所得到的圖像是相同的,即resize和通道維度的變換可交換順序
# 實(shí)際上resize只發(fā)生在spatial dimension，而通道變換發(fā)生在channels dimension，所以空間維度上的插值變換
# 是在每個(gè)通道維度上獨(dú)立進(jìn)行的。
# 另外，對(duì)于計(jì)算機(jī)而言，所讀取到的彩色圖像就是H*W*3的矩陣而已，它本身是沒有辦法區(qū)分究竟是BGR格式還是RGB格式的