使用 python 生成裁剪图像时出错

Question

我正在做一些机器学习，需要调整一些图像的大小以适应数据集的其余部分。我对图像进行了一些草率的裁剪，得到的结果让我感到困惑。 该代码非常基本，如下所示。我遇到的问题是，有时我会得到不可读且无法打开的图像，其中一些只有 40 个字节，而另一些则大小更合理。这些文件的类型为“.pgm”，并被保存为“.pgm”。

from matplotlib import pyplot
from matplotlib.image import imread
import cv2
import matplotlib.pyplot as plt
import os
import random

root = "/non-pedestrian/"
imageDir = root + "NonPedestrians/"
outDir = root + "cropped/"
xSize, ySize = (48, 96)
offset = 10 #Nothing good happends at the edges

#5000 images, 15000 crops
for idx, fname in enumerate(os.listdir(imageDir)):
    if idx >= 5000:
        break
    path = imageDir + fname
    img = cv2.imread(path, -1)
    for i in range(3):
        width, height = img.shape[:2]
        cropXmin = random.randint(offset, width - xSize - offset)
        cropYmin = random.randint(offset, height - ySize - offset)
        cropXmax = cropXmin + xSize
        cropYmax = cropYmin + ySize
        if cropYmax - cropYmin < ySize or cropXmax - cropXmin < xSize:
            # This has never happend
            print("wtf?")
            print(cropXmin, cropYmin, cropXmax, cropYmax, width, height)
            time.sleep(1000)
    
        cropped = img[cropYmin:cropYmax, cropXmin:cropXmax] # y0 : y1 , x0 : x1 
        cv2.imwrite(outDir + "{:06d}.pgm".format(idx + i * 5000), cropped)
        print("{:06d}.pgm".format(idx + i * 5000))

Answer 1

您切换了width 和height。
应该是：height, width = img.shape[:2]。

如果您不关心颜色，可以使用img = cv2.imread(path, cv2.IMREAD_GRAYSCALE)。
所有pgm 图像都必须是灰度的？

我也可以建议保护，以防输入图像太小：

if (cropXmin >= 0) and (cropYmin >= 0) and (cropXmax <= width) and (cropYmax <= height):   
    cropped = img[cropYmin:cropYmax, cropXmin:cropXmax] # y0 : y1 , x0 : x1 

---

这是一个应用单个图像和单个裁剪的代码示例：

xSize, ySize = (48, 96)

offset = 10 # Nothing good happens at the edges

# In case image is colored it's converted to Grayscale, and the crop is going to work.
img = cv2.imread(path, cv2.IMREAD_GRAYSCALE)  # Read input image as Graysacle

# height comes before width
height, width = img.shape[:2]
cropXmin = random.randint(offset, width - xSize - offset)
cropYmin = random.randint(offset, height - ySize - offset)
cropXmax = cropXmin + xSize
cropYmax = cropYmin + ySize

# Protection - required if input image is too small
if (cropXmin >= 0) and (cropYmin >= 0) and (cropXmax <= width) and (cropYmax <= height):   
    cropped = img[cropYmin:cropYmax, cropXmin:cropXmax] # y0 : y1 , x0 : x1