ImageNet 2012 - 将图像加载为数组

Question

我已经编写了代码来加载 ImageNet 2012 的片段并将其作为 np.array 返回，虽然该代码适用于大多数情况，但某些图像对于代码来说是有问题的，因为它们不是 RGB 格式。

from PIL import Image
import numpy as np
import argparse

def load_image_list (PATH, IMAGE_LIST):
    output_list = []
    for fname in IMAGE_LIST:
        img_pil = Image.open(PATH + '/' + fname)
        if img_pil.mode == 'CMYK':
            img_pil = img_pil.convert('RGB')
        width, height = img_pil.size
        smallest_dim = min((width, height))
        left = (width - smallest_dim)/2
        right = (width + smallest_dim)/2
        top = (height - smallest_dim)/2
        bottom = (height + smallest_dim)/2
        img_pil = img_pil.crop((left, top, right, bottom))
        img_pil = img_pil.resize((224,224))
        img_pil.load()

    img1 = np.asarray(img_pil, dtype="float32")
    if img1.ndim == 2:
        img1 = np.stack((img1,)*3, axis=-1)
    
    img1 = np.rollaxis(np.array(img1),2,0).astype(np.float32)
    output_list.append(img1)
try:
    output_list = np.array(output_list, dtype="float32")
except:
    counter = 0
    helper = -1
    for o in output_list:
        print(counter, o.shape)
        counter += 1
        if(o.shape[0] == 4):
            print(o)
            helper = output_list.index(o)
    counter = 0
    for name in IMAGE_LIST:
        print(counter, name)
        counter += 1


    print(helper, IMAGE_LIST[helper])
    print(helper, output_list[helper].shape)
    retry_item = IMAGE_LIST[helper]
    img_pil = Image.open(PATH + '/' + retry_item)
    width, height = img_pil.size
    smallest_dim = min((width, height))
    left = (width - smallest_dim)/2
    right = (width + smallest_dim)/2
    top = (height - smallest_dim)/2
    bottom = (height + smallest_dim)/2
    img_pil = img_pil.crop((left, top, right, bottom))
    img_pil = img_pil.resize((224,224))
    img_pil.load()

    img1 = np.asarray(img_pil, dtype="float32")
    print('number of dimensions: ', img1.ndim)
    print('shape: ', img1.shape)
    if img1.ndim == 2:
        print('ndim was 2')
        print(img1)
        img1 = np.stack((img1,)*3, axis=-1)
    img1 = np.rollaxis(np.array(img1),2,0).astype(np.float32)
    print('number of dimensions: ', img1.ndim)
    print('shape: ', img1.shape)
    
output_list = np.array(output_list, dtype="float32")  


output_list = output_list/255        
return output_list

except 行是为了找出它何时中断、为什么会中断。我已经可以加载验证集中的每个图像，但现在我尝试在测试集上运行，但图像 n02105855_2933.JPEG 再次出现错误，因为加载的图像是 4 维而不是 3 维。有没有更好的方法将未知颜色格式的 JPEG 加载为 RGB 图像？ ImageNet 2012 中还使用了哪些其他颜色格式？

Answer 1

特定图像的问题在于它是rgba格式，a尺寸均为255。所以下面的代码解决了这个问题：

from PIL import Image
import numpy as np
import argparse


def load_image_list (PATH, IMAGE_LIST):
output_list = []
for fname in IMAGE_LIST:
    img_pil = Image.open(PATH + '/' + fname)
    if img_pil.mode == 'CMYK':
        img_pil = img_pil.convert('RGB')
    width, height = img_pil.size
    smallest_dim = min((width, height))
    left = (width - smallest_dim)/2
    right = (width + smallest_dim)/2
    top = (height - smallest_dim)/2
    bottom = (height + smallest_dim)/2
    img_pil = img_pil.crop((left, top, right, bottom))
    img_pil = img_pil.resize((224,224))
    img_pil.load()

    img1 = np.asarray(img_pil, dtype="float32")
    if img1.ndim == 2:
        img1 = np.stack((img1,)*3, axis=-1)
    if img1.shape[2] == 4:
        img1 = np.delete(img1, 3, axis=2)
    
    img1 = np.rollaxis(np.array(img1),2,0).astype(np.float32)
    output_list.append(img1)
try:
    output_list = np.array(output_list, dtype="float32")
except:
    counter = 0
    helper = -1
    for o in output_list:
        print(counter, o.shape)
        counter += 1
        if(o.shape[0] == 4):
            print(o)
            helper = output_list.index(o)
    counter = 0
    for name in IMAGE_LIST:
        print(counter, name)
        counter += 1


    print(helper, IMAGE_LIST[helper])
    print(helper, output_list[helper].shape)
    retry_item = IMAGE_LIST[helper]
    img_pil = Image.open(PATH + '/' + retry_item)
    width, height = img_pil.size
    smallest_dim = min((width, height))
    left = (width - smallest_dim)/2
    right = (width + smallest_dim)/2
    top = (height - smallest_dim)/2
    bottom = (height + smallest_dim)/2
    img_pil = img_pil.crop((left, top, right, bottom))
    img_pil = img_pil.resize((224,224))
    img_pil.load()

    img1 = np.asarray(img_pil, dtype="float32")
    print('number of dimensions: ', img1.ndim)
    print('shape: ', img1.shape)
    if img1.ndim == 2:
        print('ndim was 2')
        print(img1)
        img1 = np.stack((img1,)*3, axis=-1)
    if img1.shape[2] == 4:
        print('color dimension was size 4')
        img1 = np.delete(img1, 3, axis=2)
    
    img1 = np.rollaxis(np.array(img1),2,0).astype(np.float32)
    print('number of dimensions: ', img1.ndim)
    print('shape: ', img1.shape)
    
    output_list = np.array(output_list, dtype="float32")  


output_list = output_list/255        
return output_list