图像处理

PILLOW

由PIL发展而来，PIL已经不更新了

pillow-simd

推荐安装avx2指令集，更快，查看cpu是否支持avx2指令集：
gcc -mavx2 -dM -E - < /dev/null | egrep “SSE|AVX” | sort

SCIKIT-IMAGE

由numpy和scipy发展而来，针对numpy数组操作

OPENCV

计算机视觉库，更底层

速度测试

测试pillow-simd，scikit-image，torchvision.transform(用的pillow)时间：

begin = time.time()
for i in range(100):
    a = Image.open(img)
    a = a.resize((120,160))
    a = a.crop((0,0,100,100))
    a = np.array(a, dtype=float)
    a -= [100,100,100]
    a /= [1,1,1]
    a/=255
pil = time.time()
for j in range(100):
    a = skimage.data.imread(img)
    a = skimage.transform.resize(a, (120,160))
    a = a[0:100,0:100,:]
    a -= [100,100,100]
    a /= [1,1,1]
    a/=255
ski = time.time()
for k in range(100):
    transform = transforms.Compose([
        transforms.Scale((120,160)),
        transforms.RandomSizedCrop(100),
        transforms.ToTensor(),
        transforms.Normalize((100/255,100/255,100/255), (0.2,0.2,0.2))
    ])
    a = Image.open(img)
    a = transform(a)
tra = time.time()
print(pil-begin)
print(ski-pil)
print(tra-ski)

结果

0.15108823776245117
0.5605413913726807
0.24411320686340332

推荐方案：

a = Image.open(img)
a = a.resize((120, 160), Image.NEAREST)
a = a.crop((0, 0, 100, 100))
a = np.array(a, dtype=float)
a = (a-[100, 100, 100])/[255, 255, 255]