實習中的小技能-CV算法篇-2

1. Pytorch2Caffe

Pytorch2caffe
example路徑下可直接轉，如有全局平均池化，需自己在prototxt中加入，測試squeezenet沒有問題

2. 數據增強-imgaug

import random
from imgaug import augmenters as iaa
from imgaug import parameters as iap
from itertools import chain
import imgaug as ia
import cv2
import imageio as imgio
import numpy as np


def _arithmetic(num):
    supported_ops = [
        iaa.Add((-40, 40)),
        iaa.AddElementwise((-40, 40), per_channel=0.5),
        iaa.AdditiveGaussianNoise((0, 0.2 * 255), per_channel=0.5),
        iaa.AdditiveLaplaceNoise(scale=0.2*255, per_channel=True),
        iaa.AdditivePoissonNoise((0, 40), per_channel=True),
        iaa.Multiply((0.5, 1.5), per_channel=0.5),
        iaa.MultiplyElementwise((0.5, 1.5), per_channel=0.5),
        iaa.Dropout(p=(0, 0.2), per_channel=0.5),
        iaa.CoarseDropout((0.0, 0.05), size_percent=(0.02, 0.25)),
        iaa.CoarseDropout(0.02, size_percent=0.15, per_channel=0.5),
        iaa.ReplaceElementwise(0.1, iap.Normal(128, 0.4*128), per_channel=0.5),
        iaa.ReplaceElementwise(
            iap.FromLowerResolution(iap.Binomial(0.1), size_px=8),
            iap.Normal(128, 0.4*128),
            per_channel=0.5),
        iaa.ImpulseNoise(0.1),
        iaa.SaltAndPepper(0.1, per_channel=True),
        iaa.CoarseSaltAndPepper(0.05, size_percent=(0.01, 0.1)),
        iaa.CoarseSaltAndPepper(
            0.05, size_percent=(0.01, 0.1), per_channel=True),
        iaa.Salt(0.1),
        iaa.CoarseSalt(0.05, size_percent=(0.01, 0.1)),
        iaa.Pepper(0.1),
        iaa.CoarsePepper(0.05, size_percent=(0.01, 0.1)),
        iaa.Invert(0.5),
        iaa.Invert(0.25, per_channel=0.5),
        iaa.ContrastNormalization((0.5, 1.5)),
        iaa.ContrastNormalization((0.5, 1.5), per_channel=0.5),
        iaa.JpegCompression(compression=(70, 99))
    ]
    if num > len(supported_ops):
        return "the number must less %s" % len(supported_ops)
    return random.sample(supported_ops, num)


def _blend(num):
    supported_ops = [
        iaa.Alpha(0.5, iaa.Grayscale(1.0)),
        iaa.Alpha((0.0, 1.0), iaa.Grayscale(1.0)),
        iaa.Alpha(
            (0.0, 1.0),
            iaa.Affine(rotate=(-20, 20)),
            per_channel=0.5),
        iaa.Alpha(
            (0.0, 1.0),
            first=iaa.Add(100),
            second=iaa.Multiply(0.2)),
        iaa.Alpha([0.25, 0.75], iaa.MedianBlur(13)),
        iaa.AlphaElementwise(0.5, iaa.Grayscale(1.0)),
        iaa.AlphaElementwise((0, 1.0), iaa.Grayscale(1.0)),
        iaa.AlphaElementwise(
            (0.0, 1.0),
            iaa.Affine(rotate=(-20, 20)),
            per_channel=0.5),
        iaa.AlphaElementwise(
            (0.0, 1.0),
            first=iaa.Add(100),
            second=iaa.Multiply(0.2)),
        iaa.AlphaElementwise([0.25, 0.75], iaa.MedianBlur(13)),
        iaa.SimplexNoiseAlpha(
            iaa.EdgeDetect(1.0),
            upscale_method="nearest"),
        iaa.SimplexNoiseAlpha(
            iaa.EdgeDetect(1.0),
            sigmoid_thresh=iap.Normal(10.0, 5.0)),
        iaa.FrequencyNoiseAlpha(first=iaa.EdgeDetect(1.0)),
        iaa.FrequencyNoiseAlpha(
            first=iaa.EdgeDetect(1.0),
            upscale_method="nearest"),
        iaa.FrequencyNoiseAlpha(
            first=iaa.EdgeDetect(1.0),
            upscale_method="linear"),
        iaa.FrequencyNoiseAlpha(
            first=iaa.EdgeDetect(1.0),
            upscale_method="linear",
            exponent=-2,
            sigmoid=False),
        iaa.FrequencyNoiseAlpha(
            first=iaa.EdgeDetect(1.0),
            sigmoid_thresh=iap.Normal(10.0, 5.0))
    ]
    if num > len(supported_ops):
        return "the number must less %s" % len(supported_ops)
    return random.sample(supported_ops, num)


def _blur(num):
    supported_ops = [
        iaa.GaussianBlur(sigma=(0.0, 3.0)),
        iaa.AverageBlur(k=(2, 11)),
        iaa.AverageBlur(k=((5, 11), (1, 3))),
        iaa.MedianBlur(k=(3, 11)),
        iaa.BilateralBlur(
            d=(3, 10), sigma_color=(10, 250), sigma_space=(10, 250)),
        iaa.MotionBlur(k=15),
        iaa.MotionBlur(k=15, angle=[-45, 45]),
    ]
    if num > len(supported_ops):
        return "the number must less %s" % len(supported_ops)
    return random.sample(supported_ops, num)


def _color(num):
    supported_ops = [
        iaa.WithColorspace(
            to_colorspace="HSV",
            from_colorspace="RGB",
            children=iaa.WithChannels(
                0,
                iaa.Add((0, 50))
            )
        ),
        iaa.WithHueAndSaturation(
            iaa.WithChannels(0, iaa.Add((0, 50)))
        ),
        iaa.WithHueAndSaturation([
            iaa.WithChannels(0, iaa.Add((-30, 10))),
            iaa.WithChannels(1, [
                iaa.Multiply((0.5, 1.5)),
                iaa.LinearContrast((0.75, 1.25))
            ])
        ]),
        iaa.MultiplyHueAndSaturation((0.5, 1.5), per_channel=True),
        iaa.MultiplyHueAndSaturation(mul_hue=(0.5, 1.5)),
        iaa.MultiplyHueAndSaturation(mul_saturation=(0.5, 1.5)),
        iaa.MultiplyHue((0.5, 1.5)),
        iaa.MultiplySaturation((0.5, 1.5)),
        iaa.AddToHueAndSaturation((-50, 50), per_channel=True),
        iaa.AddToHue((-50, 50)),
        iaa.AddToSaturation((-50, 50)),
        iaa.Sequential([
            iaa.ChangeColorspace(from_colorspace="RGB", to_colorspace="HSV"),
            iaa.WithChannels(0, iaa.Add((50, 100))),
            iaa.ChangeColorspace(from_colorspace="HSV", to_colorspace="RGB")
        ]),
        iaa.Grayscale(alpha=(0.0, 1.0)),
        iaa.KMeansColorQuantization(),
        iaa.KMeansColorQuantization(n_colors=(4, 16)),
        iaa.KMeansColorQuantization(
            to_colorspace=[iaa.ChangeColorspace.RGB, iaa.ChangeColorspace.HSV]),
        iaa.UniformColorQuantization(),
        iaa.UniformColorQuantization(n_colors=8),
        iaa.UniformColorQuantization(n_colors=(4, 16)),
        iaa.UniformColorQuantization(
            from_colorspace=iaa.ChangeColorspace.BGR,
            to_colorspace=[iaa.ChangeColorspace.RGB, iaa.ChangeColorspace.HSV])
    ]
    if num > len(supported_ops):
        return "the number must less %s" % len(supported_ops)
    return random.sample(supported_ops, num)


def _contrast(num):
    supported_ops = [
        iaa.GammaContrast((0.5, 2.0)),
        iaa.GammaContrast((0.5, 2.0), per_channel=True),
        iaa.SigmoidContrast(gain=(3, 10), cutoff=(0.4, 0.6)),
        iaa.SigmoidContrast(
            gain=(3, 10), cutoff=(0.4, 0.6), per_channel=True),
        iaa.LogContrast(gain=(0.6, 1.4)),
        iaa.LogContrast(gain=(0.6, 1.4), per_channel=True),
        iaa.LinearContrast((0.4, 1.6)),
        iaa.LinearContrast((0.4, 1.6), per_channel=True),
        iaa.AllChannelsCLAHE(),
        iaa.AllChannelsCLAHE(clip_limit=(1, 10)),
        iaa.AllChannelsCLAHE(clip_limit=(1, 10), per_channel=True),
        iaa.CLAHE(
            tile_grid_size_px=iap.Discretize(iap.Normal(loc=7, scale=2)),
            tile_grid_size_px_min=3),
        iaa.CLAHE(
            from_colorspace=iaa.CLAHE.BGR,
            to_colorspace=iaa.CLAHE.HSV),
        iaa.AllChannelsHistogramEqualization(),
        iaa.Alpha((0.0, 1.0), iaa.AllChannelsHistogramEqualization()),
        iaa.HistogramEqualization(
            from_colorspace=iaa.HistogramEqualization.BGR,
            to_colorspace=iaa.HistogramEqualization.HSV)
    ]
    if num > len(supported_ops):
        return "the number must less %s" % len(supported_ops)
    return random.sample(supported_ops, num)


def _flip(num):
    supported_ops = [
        iaa.Fliplr(0.5),
        iaa.Flipud(0.5)
    ]
    if num > len(supported_ops):
        return "the number must less %s" % len(supported_ops)
    return random.sample(supported_ops, num)

def _geometric(num):
    supported_ops = [
        iaa.Affine(scale=(0.5, 1.5)),
        iaa.Affine(scale={"x": (0.5, 1.5), "y": (0.5, 1.5)}),
        iaa.Affine(translate_percent={"x": (-0.2, 0.2), "y": (-0.2, 0.2)}),
        iaa.Affine(rotate=(-45, 45)),
        iaa.Affine(translate_percent={"x": -0.20}, mode=ia.ALL, cval=(0, 255)),
        iaa.PerspectiveTransform(scale=(0.01, 0.15)),
        iaa.ElasticTransformation(alpha=(0, 5.0), sigma=0.25)
    ]
    if num > len(supported_ops):
        return "the number must less %s" % len(supported_ops)
    return random.sample(supported_ops, num)

def _pooling(num):
    supported_ops = [
        iaa.AveragePooling([2, 8]),
        iaa.MaxPooling(2),
        iaa.MaxPooling([2, 8]),
        iaa.MaxPooling(((1, 7), (1, 7))),
        iaa.MinPooling([2, 8]),
        iaa.MinPooling(((1, 7), (1, 7)))
    ]
    if num > len(supported_ops):
        return "the number must less %s" % len(supported_ops)
    return random.sample(supported_ops, num)


def _weather(num):
    supported_ops = [
        iaa.FastSnowyLandscape(
            lightness_threshold=140,
            lightness_multiplier=2.5
        ),
        iaa.FastSnowyLandscape(
            lightness_threshold=[128, 200],
            lightness_multiplier=(1.5, 3.5)
        ),
        iaa.Clouds(),
        iaa.Snowflakes(flake_size=(0.7, 0.95), speed=(0.001, 0.03))
    ]
    if num > len(supported_ops):
        return "the number must less %s" % len(supported_ops)
    return random.sample(supported_ops, num)
class img_aug(object):

    def __init__(self, all_some_one='all', use_arithmetic_num=None, use_blend_num=None, use_blur_num=None,
                 use_color_num=None, use_contrast_num=None, use_flip_num=None, use_geometric_num=None,
                 use_pooling_num=None, use_weather_num=None):

        self.flag = all_some_one
        self.use_arithmetic_num = use_arithmetic_num
        self.use_blend_num = use_blend_num
        self.use_blur_num = use_blur_num
        self.use_color_num = use_color_num
        self.use_contrast_num = use_contrast_num
        self.use_flip_num = use_flip_num
        self.use_geometric_num = use_geometric_num
        self.use_pooling_num = use_pooling_num
        self.use_weather_num = use_weather_num
        self.functions = []
        if self.use_arithmetic_num:
            self.functions.append(_arithmetic(self.use_arithmetic_num))
        if self.use_blend_num:
            self.functions.append(_arithmetic(self.use_blend_num))
        if self.use_blur_num:
            self.functions.append(_arithmetic(self.use_blur_num))
        if self.use_color_num:
            self.functions.append(_arithmetic(self.use_color_num))
        if self.use_contrast_num:
            self.functions.append(_arithmetic(self.use_contrast_num))
        if self.use_flip_num:
            self.functions.append(_arithmetic(self.use_flip_num))
        if self.use_pooling_num:
            self.functions.append(_arithmetic(self.use_pooling_num))
        if self.use_arithmetic_num:
            self.functions.append(_arithmetic(self.use_arithmetic_num))
        if self.use_weather_num:
            self.functions.append(_arithmetic(self.use_weather_num))

        self.functions = list(chain.from_iterable(self.functions))

    def __call__(self, img):

        if self.flag == 'all':
            return iaa.Sequential(self.functions).augment_images(img)
        elif self.flag == 'some':
            return iaa.SomeOf((0, len(self.functions)), self.functions).augment_images(img)
        elif self.flag == 'one':
            return iaa.OneOf(self.functions).augment_images(img)


if __name__ == "__main__":
    imgfile_add = 'xxx.jpg'
    imgfile_clear = 'xxx.jpg'
    imgfile_push = 'xxx.jpg'
    add = cv2.imread(imgfile_add)
    push = cv2.imread(imgfile_push)
    clear = cv2.imread(imgfile_clear)

    b, g, r = cv2.split(add)
    add = cv2.merge([r, g, b])

    b, g, r = cv2.split(push)
    push = cv2.merge([r, g, b])

    b, g, r = cv2.split(clear)
    clear = cv2.merge([r, g, b])

    imgs =[add, push, clear]
    img_aug = img_aug(all_some_one='one', use_arithmetic_num=3, use_weather_num=1, use_blur_num=2,  use_blend_num=1, use_contrast_num=0, use_flip_num=1, use_geometric_num=1)

    ia.imshow(np.hstack(img_aug(imgs)))
    # ia.imshow(np.hstack(imgs))

3. 圖片歸一化

3.1 Pytorch

import torchvision.transforms as transforms
import numpy as np
from PIL import Image
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                      std=[0.229, 0.224, 0.225])
#transforms.ToTensor()
transform1 = transforms.Compose([
    transforms.Resize([224, 224]),
    transforms.ToTensor(),
    normalize,
    ]
 )
# ##numpy.ndarray
imgfile= 'xxx.jpg'
img = Image.open(imgfile).convert('RGB')# 讀取圖像
#print(img)
#img = cv2.resize(img,(224,244))
img = transform1(img) # 歸一化到 [0.0,1.0]
# img = img.unsqueeze(0)
img

3.2 PIL

import torchvision.transforms as transforms
import numpy as np
from PIL import Image

imgfile= 'xxx.jpg'
img = Image.open(imgfile).convert('RGB')# 讀取圖像
#print(img)
img = img.resize((224, 224), Image.BILINEAR)
#img = cv2.resize(img,(224,244))
# img = np.array(img)/255
img = img.transpose(2, 0, 1)
# mu = np.array([[[0.485]],[[0.456]],[[0.406]]])
# sigma = np.array([[[0.229]],[[0.224]],[[0.225]]])
# img = (img - mu)/sigma
img = np.around(img, decimals=4)
img


img2 = cv2.imread(imgfile)
b,g,r = cv2.split(img2)
img2 = cv2.merge([r,g,b])
img2 = cv2.resize(img2, (224, 224), interpolation=cv2.INTER_LINEAR)  

img2 = np.array(img2)/255
img2 = img2.transpose(2, 0, 1)
mu = np.array([[[0.485]],[[0.456]],[[0.406]]])
sigma = np.array([[[0.229]],[[0.224]],[[0.225]]])
img2 = (img2 - mu)/sigma
img2 = np.around(img2, decimals=4)
img2

diff = abs(img - img2)
diff = np.mean(diff)
diff

3.3 OpenCV

import cv2
import numpy as np
imgfile= 'xxx.jpg'
img = cv2.imread(imgfile)
b,g,r = cv2.split(img)
img = cv2.merge([r,g,b])
img = cv2.resize(img, (224, 224), interpolation=cv2.INTER_LINEAR)  
cv2.imwrite('xxxt.jpg', img)
img = np.array(img)/255
img = img.transpose(2, 0, 1)
mu = np.array([[[0.485]],[[0.456]],[[0.406]]])
sigma = np.array([[[0.229]],[[0.224]],[[0.225]]])
img = (img - mu)/sigma
img = np.around(img, decimals=4)
img