数据增强(Data Augmentation)

这篇具有很好参考价值的文章主要介绍了数据增强(Data Augmentation)。希望对大家有所帮助。如果存在错误或未考虑完全的地方,请大家不吝赐教,您也可以点击"举报违法"按钮提交疑问。


前言

最近写论文需要插入很多图片,为了蒙混过关,找了很多很多数据增强的手段,增强论文的丰富性,大家不要学我哈,反正我把技巧放这儿了!!!哈哈哈哈哈哈哈哈哈
数据增强之后效果变差,激光条纹中心提取,python,python,计算机视觉,数据增强

1. Data Augmentation

1.1 数据增强的作用

大家都知道在深度学习网络训练中,模型的样本越充足训练出来的网络模型泛化性越强,鲁棒性越高。最好的例子就是SSD对大目标效果很好,对小目标效果很差,但当使用数据增强后,涨分立马上去了,原因就在于数据增强crop,可以让小目标变成大目标。

  1. 增加训练的数据量,提高模型的泛化能力
  2. 增加噪声数据,提升模型的鲁棒性
  3. 一定程度上能解决过拟合问题,样本过少容易出现过拟合
  4. 解决样本不平衡问题,例如某个类别过少,数据增强可以增强这个类别的数量。

1.2 图像增强小工具

我写了一款图像增强的小工具,大家可以查看连接下载使用
数据增强之后效果变差,激光条纹中心提取,python,python,计算机视觉,数据增强

2. 数据增强的手段

数据增强的手段太多了,我就不一一解释了,直接总结吧

  1. 几何变换:反转/平移/旋转/缩放/裁剪
  2. 颜色变换:加噪声/亮度/饱和度/颜色反转/直方图均衡/色彩平衡/
  3. 基于GAN网络数据增强:用生成模型,生成不同风格的图像
  4. 高级增强:Cutout/Mixup/One-hot/Cutmix/AutoAugment/Sample pairing/FMix
  5. 网络中增强:RandAugment/Refinement/Mosaic/DropOut/DropOut/DropBlock

下面简单是介绍一些大家可能不太熟悉的概念:
Cutout:图像中挖掉一个矩形快补0
Mixup:image——两张图片线性叠加,label——one-hot形式进行相同的线性叠加
Cutmix:Cutout + Mixup,image——在图片中挖出来的矩形用另外一张图片填充,label——做同样的叠加
Sample pairing:image——两张图片直接做平均,label——选取其中一张
FMix:采用Fourier变换选取挖掉和补充的区域
推荐一下网站:
https://segmentfault.com/a/1190000022784467
https://jishuin.proginn.com/p/763bfbd29660
https://www.pythonf.cn/read/169722

3. 数据增强的代码

3.1 代码

如果大家想深入了解,建议大家跑跑代码

import os
import cv2
import numpy as np

class Resize(object):
    """
    调整大小
    """

    def __init__(self, output_size):
        self.output_size = output_size

    def __call__(self, X, Y):
        _X = cv2.resize(X, self.output_size, interpolation=cv2.INTER_NEAREST)
        w, h = self.output_size
        c = Y.shape[-1]
        # _Y = cv2.resize(Y, self.output_size)
        _Y = np.zeros((h, w, c))
        for i in range(Y.shape[-1]):
            _Y[..., i] = cv2.resize(Y[..., i], self.output_size, interpolation=cv2.INTER_NEAREST)
        return _X, _Y[...,0]

class Clip(object):
    """
    彩色截断
    """
    def __init__(self, mini, maxi=None):
        if maxi is None:
            self.mini, self.maxi = 0, mini
        else:
            self.mini, self.maxi = mini, maxi

    def __call__(self, X, Y):
        mini_mask = np.where(X < self.mini)
        maxi_mask = np.where(X > self.maxi)
        X[mini_mask] = self.mini
        X[maxi_mask] = self.maxi
        return X, Y

class Normalize(object):
    """
   最大最小值归一化
    """
    def __init__(self, axis=None):
        self.axis = axis

    def __call__(self, X, Y):
        mini = np.min(X, self.axis)
        maxi = np.max(X, self.axis)
        X = (X - mini) / (maxi - mini)
        X = 255 * X
        return X, Y

class Standardize(object):
    """
    标准归一化
    """

    def __init__(self, axis=None):
        self.axis = axis

    def __call__(self, X, Y):
        mean =  np.mean(X, self.axis)
        std = np.std(X, self.axis)
        X = (X - mean) / std
        X = 255 * X
        return X, Y

class Flip(object):
    """
    反转
    """
    def __call__(self, X, Y):
        for axis in [0, 1]:
            if np.random.rand(1) < 0.5:
                X = np.flip(X, axis)
                Y = np.flip(Y, axis)
        return X, Y

class Crop(object):
    def __init__(self, min_size_ratio, max_size_ratio=(1, 1)):
        self.min_size_ratio = np.array(list(min_size_ratio))
        self.max_size_ratio = np.array(list(max_size_ratio))

    def __call__(self, X, Y):
        size = np.array(X.shape[:2])
        mini = self.min_size_ratio * size
        maxi = self.max_size_ratio * size
        # random size
        h = np.random.randint(mini[0], maxi[0])
        w = np.random.randint(mini[1], maxi[1])
        # random place
        shift_h = np.random.randint(0, size[0] - h)
        shift_w = np.random.randint(0, size[1] - w)
        X = X[shift_h:shift_h+h, shift_w:shift_w+w]
        Y = Y[shift_h:shift_h+h, shift_w:shift_w+w]

        return X, Y

class CustomFilter(object):
    def __init__(self, kernel):
        self.kernel = kernel

    def __call__(self, X, Y):
        X = cv2.filter2D(X, -1, self.kernel)
        return X, Y

class Sharpen(object):
    """
    锐化
    """
    def __init__(self, max_center=4):
        self.max_center = max_center
        self.identity = np.array([[0, 0, 0],
                                  [0, 1, 0],
                                  [0, 0, 0]])
        self.sharpen = np.array([[ 0, -1,  0],
                                [-1,  4, -1],
                                [ 0, -1,  0]]) / 4

    def __call__(self, X, Y):

        sharp = self.sharpen * np.random.random() * self.max_center
        kernel = self.identity + sharp

        X = cv2.filter2D(X, -1, kernel)
        return X, Y

class GaussianBlur(object):
    """
    高斯滤波
    """
    def __init__(self, max_kernel=np.array([7, 7])):
        self.max_kernel = ((max_kernel + 1) // 2)

    def __call__(self, X, Y):
        kernel_size = (
            np.random.randint(1, self.max_kernel[0]) * 2 + 1,
            np.random.randint(1, self.max_kernel[1]) * 2 + 1,
        )
        X = cv2.GaussianBlur(X, kernel_size, 0)
        return X, Y

class Perspective(object):
    """
    透视
    """
    def __init__(self,
                 max_ratio_translation=(0.2, 0.2, 0),
                 max_rotation=(10, 10, 360),
                 max_scale=(0.1, 0.1, 0.2),
                 max_shearing=(15, 15, 5)):

        self.max_ratio_translation = np.array(max_ratio_translation)
        self.max_rotation = np.array(max_rotation)
        self.max_scale = np.array(max_scale)
        self.max_shearing = np.array(max_shearing)

    def __call__(self, X, Y):

        # get the height and the width of the image
        h, w = X.shape[:2]
        max_translation = self.max_ratio_translation * np.array([w, h, 1])
        # get the values on each axis
        t_x, t_y, t_z = np.random.uniform(-1, 1, 3) * max_translation
        r_x, r_y, r_z = np.random.uniform(-1, 1, 3) * self.max_rotation
        sc_x, sc_y, sc_z = np.random.uniform(-1, 1, 3) * self.max_scale + 1
        sh_x, sh_y, sh_z = np.random.uniform(-1, 1, 3) * self.max_shearing

        # convert degree angles to rad
        theta_rx = np.deg2rad(r_x)
        theta_ry = np.deg2rad(r_y)
        theta_rz = np.deg2rad(r_z)
        theta_shx = np.deg2rad(sh_x)
        theta_shy = np.deg2rad(sh_y)
        theta_shz = np.deg2rad(sh_z)


        # compute its diagonal
        diag = (h ** 2 + w ** 2) ** 0.5
        # compute the focal length
        f = diag
        if np.sin(theta_rz) != 0:
            f /= 2 * np.sin(theta_rz)

        # set the image from cartesian to projective dimension
        H_M = np.array([[1, 0, -w / 2],
                        [0, 1, -h / 2],
                        [0, 0,      1],
                        [0, 0,      1]])
        # set the image projective to carrtesian dimension
        Hp_M = np.array([[f, 0, w / 2, 0],
                         [0, f, h / 2, 0],
                         [0, 0,     1, 0]])

        # adjust the translation on z
        t_z = (f - t_z) / sc_z ** 2
        # translation matrix to translate the image
        T_M = np.array([[1, 0, 0, t_x],
                        [0, 1, 0, t_y],
                        [0, 0, 1, t_z],
                        [0, 0, 0,  1]])

        # calculate cos and sin of angles
        sin_rx, cos_rx = np.sin(theta_rx), np.cos(theta_rx)
        sin_ry, cos_ry = np.sin(theta_ry), np.cos(theta_ry)
        sin_rz, cos_rz = np.sin(theta_rz), np.cos(theta_rz)
        # get the rotation matrix on x axis
        R_Mx = np.array([[1,      0,       0, 0],
                         [0, cos_rx, -sin_rx, 0],
                         [0, sin_rx,  cos_rx, 0],
                         [0,      0,       0, 1]])
        # get the rotation matrix on y axis
        R_My = np.array([[cos_ry, 0, -sin_ry, 0],
                         [     0, 1,       0, 0],
                         [sin_ry, 0,  cos_ry, 0],
                         [     0, 0,       0, 1]])
        # get the rotation matrix on z axis
        R_Mz = np.array([[cos_rz, -sin_rz, 0, 0],
                         [sin_rz,  cos_rz, 0, 0],
                         [     0,       0, 1, 0],
                         [     0,       0, 0, 1]])
        # compute the full rotation matrix
        R_M = np.dot(np.dot(R_Mx, R_My), R_Mz)

        # get the scaling matrix
        Sc_M = np.array([[sc_x,     0,    0, 0],
                         [   0,  sc_y,    0, 0],
                         [   0,     0, sc_z, 0],
                         [   0,     0,    0, 1]])

        # get the tan of angles
        tan_shx = np.tan(theta_shx)
        tan_shy = np.tan(theta_shy)
        tan_shz = np.tan(theta_shz)
        # get the shearing matrix on x axis
        Sh_Mx = np.array([[      1, 0, 0, 0],
                          [tan_shy, 1, 0, 0],
                          [tan_shz, 0, 1, 0],
                          [      0, 0, 0, 1]])
        # get the shearing matrix on y axis
        Sh_My = np.array([[1, tan_shx, 0, 0],
                          [0,       1, 0, 0],
                          [0, tan_shz, 1, 0],
                          [0,       0, 0, 1]])
        # get the shearing matrix on z axis
        Sh_Mz = np.array([[1, 0, tan_shx, 0],
                          [0, 1, tan_shy, 0],
                          [0, 0,       1, 0],
                          [0, 0,       0, 1]])
        # compute the full shearing matrix
        Sh_M = np.dot(np.dot(Sh_Mx, Sh_My), Sh_Mz)

        Identity = np.array([[1, 0, 0, 0],
                             [0, 1, 0, 0],
                             [0, 0, 1, 0],
                             [0, 0, 0, 1]])

        # compute the full transform matrix
        M = Identity
        M = np.dot(Sh_M, M)
        M = np.dot(R_M,  M)
        M = np.dot(Sc_M, M)
        M = np.dot(T_M,  M)
        M = np.dot(Hp_M, np.dot(M, H_M))
        # apply the transformation
        X = cv2.warpPerspective(X, M, (w, h))
        Y = cv2.warpPerspective(Y, M, (w, h))
        return X, Y

class Cutout(object):
    """
    随机擦除
    """
    def __init__(self,
                 min_size_ratio,
                 max_size_ratio,
                 channel_wise=False,
                 crop_target=True,
                 max_crop=10,
                 replacement=0):
        self.min_size_ratio = np.array(list(min_size_ratio))
        self.max_size_ratio = np.array(list(max_size_ratio))
        self.channel_wise = channel_wise
        self.crop_target = crop_target
        self.max_crop = max_crop
        self.replacement = replacement

    def __call__(self, X, Y):
        size = np.array(X.shape[:2])
        mini = self.min_size_ratio * size
        maxi = self.max_size_ratio * size
        for _ in range(self.max_crop):
            # random size
            h = np.random.randint(mini[0], maxi[0])
            w = np.random.randint(mini[1], maxi[1])
            # random place
            shift_h = np.random.randint(0, size[0] - h)
            shift_w = np.random.randint(0, size[1] - w)
            if self.channel_wise:
                c = np.random.randint(0, X.shape[-1])
                X[shift_h:shift_h+h, shift_w:shift_w+w, c] = self.replacement
                if self.crop_target:
                    Y[shift_h:shift_h+h, shift_w:shift_w+w] = self.replacement
            else:
                X[shift_h:shift_h+h, shift_w:shift_w+w] = self.replacement
                if self.crop_target:
                    Y[shift_h:shift_h+h, shift_w:shift_w+w] = self.replacement
        return X, Y

class Leaf(object):
    def __init__(self):
        pass
    def __call__(self, X, Y):
        blur = cv2.GaussianBlur(X, (7, 7), 0)
        hsv_blur = cv2.cvtColor(blur, cv2.COLOR_BGR2HSV)
        # lower mask (0-10)
        lower_red = np.array([0,130,130])
        upper_red = np.array([20,255,255])
        mask_0 = cv2.inRange(hsv_blur, lower_red, upper_red)

        # upper mask (170-180)
        lower_red = np.array([165,130,130])
        upper_red = np.array([185,255,255])
        mask_1 = cv2.inRange(hsv_blur, lower_red, upper_red)
        hsv_blur[np.where(mask_1)] = hsv_blur[np.where(mask_1)] - np.array([165, 0, 0])

        mask = mask_0 + mask_1
        # change color
        turn_color = np.random.randint(0, 255)
        hsv_blur[np.where(mask)] = hsv_blur[np.where(mask)] + np.array([turn_color, 0, 0])
        X_blur = cv2.cvtColor(hsv_blur, cv2.COLOR_HSV2BGR)
        X[np.where(mask)] = X_blur[np.where(mask)]
        return X, Y

class Brightness(object):
    """
    随机改变图像的亮度
    """
    def __init__(self, range_brightness=(-50, 50)):
        self.range_brightness = range_brightness

    def __call__(self, X, Y):
        brightness = np.random.randint(*self.range_brightness)
        X = X + brightness
        return X, Y

class Contrast(object):
    """
    改变图像的对比度
    """

    def __init__(self, range_contrast=(-50, 50)):
        self.range_contrast = range_contrast

    def __call__(self, X, Y):
        contrast = np.random.randint(*self.range_contrast)
        X = X * (contrast / 127 + 1) - contrast
        return X, Y

class UniformNoise(object):
    """
    均匀噪声
    """
    def __init__(self, low=-50, high=50):
        self.low = low
        self.high = high

    def __call__(self, X, Y):
        noise = np.random.uniform(self.low, self.high, X.shape)
        X = X + noise
        return X, Y

class GaussianNoise(object):
    def __init__(self, center=0, std=50):
        self.center = center
        self.std = std

    def __call__(self, X, Y):
        noise = np.random.normal(self.center, self.std, X.shape)
        X = X + noise
        return X, Y

class Vignetting(object):
    def __init__(self,
                 ratio_min_dist=0.2,
                 range_vignette=(0.2, 0.8),
                 random_sign=False):
        self.ratio_min_dist = ratio_min_dist
        self.range_vignette = np.array(range_vignette)
        self.random_sign = random_sign

    def __call__(self, X, Y):
        h, w = X.shape[:2]
        min_dist = np.array([h, w]) / 2 * np.random.random() * self.ratio_min_dist

        # create matrix of distance from the center on the two axis
        x, y = np.meshgrid(np.linspace(-w/2, w/2, w), np.linspace(-h/2, h/2, h))
        x, y = np.abs(x), np.abs(y)

        # create the vignette mask on the two axis
        x = (x - min_dist[0]) / (np.max(x) - min_dist[0])
        x = np.clip(x, 0, 1)
        y = (y - min_dist[1]) / (np.max(y) - min_dist[1])
        y = np.clip(y, 0, 1)

        # then get a random intensity of the vignette
        vignette = (x + y) / 2 * np.random.uniform(*self.range_vignette)
        vignette = np.tile(vignette[..., None], [1, 1, 3])

        sign = 2 * (np.random.random() < 0.5) * (self.random_sign) - 1
        X = X * (1 + sign * vignette)

        return X, Y

class LensDistortion(object):
    def __init__(self, d_coef=(0.15, 0.15, 0.1, 0.1, 0.05)):
        self.d_coef = np.array(d_coef)

    def __call__(self, X, Y):

        # get the height and the width of the image
        h, w = X.shape[:2]

        # compute its diagonal
        f = (h ** 2 + w ** 2) ** 0.5

        # set the image projective to carrtesian dimension
        K = np.array([[f, 0, w / 2],
                      [0, f, h / 2],
                      [0, 0,     1]])

        d_coef = self.d_coef * np.random.random(5) # value
        d_coef = d_coef * (2 * (np.random.random(5) < 0.5) - 1) # sign
        # Generate new camera matrix from parameters
        M, _ = cv2.getOptimalNewCameraMatrix(K, d_coef, (w, h), 0)

        # Generate look-up tables for remapping the camera image
        remap = cv2.initUndistortRectifyMap(K, d_coef, None, M, (w, h), 5)

        # Remap the original image to a new image
        X = cv2.remap(X, *remap, cv2.INTER_LINEAR)
        Y = cv2.remap(Y, *remap, cv2.INTER_LINEAR)
        return X, Y


def data_aug(img_path,  label_path, save_img, save_label):
    imgs = os.listdir(img_path)
    for im_name in imgs:
        name = im_name.split('.')[0]
        im_full_path = os.path.join(img_path, im_name)
        lab_full_path = os.path.join(label_path, im_name)

        
        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path)
        # save_img_path = os.path.join(save_img, name + '_Resize960.png')
        # save_lab_path = os.path.join(save_label, name + '_Resize960.png')
        # im_Resize, lab_Resize = Resize((960,960))(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im_Resize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab_Resize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path)
        # save_img_path = os.path.join(save_img, name + '_Resize384.png')
        # save_lab_path = os.path.join(save_label, name + '_Resize384.png')
        # im_Resize, lab_Resize = Resize((384,384))(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im_Resize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab_Resize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_Clip.png')
        # save_lab_path = os.path.join(save_label, name + '_Clip.png')
        # im_Clip, lab_Clip = Clip(20,150)(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im_Clip, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab_Clip, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_Normalize.png')
        # save_lab_path = os.path.join(save_label, name + '_Normalize.png')
        # im_Normalize, lab_Normalize = Normalize(1)(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im_Normalize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab_Normalize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_Standardize.png')
        # save_lab_path = os.path.join(save_label, name + '_Standardize.png')
        # im_Standardize, lab_Standardize = Standardize(1)(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im_Standardize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab_Standardize, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path)
        # save_img_path = os.path.join(save_img, name + '_Flip.png')
        # save_lab_path = os.path.join(save_label, name + '_Flip.png')
        # im__Flip, lab__Flip= Flip()(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im__Flip, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab__Flip, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_Crop.png')
        # save_lab_path = os.path.join(save_label, name + '_Crop.png')
        # im__Crop, lab__Crop = Crop((0.5,0.5))(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im__Crop, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab__Crop, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_CustomFilter.png')
        # save_lab_path = os.path.join(save_label, name + '_CustomFilter.png')
        # lpls = np.array([[-1,1,-1],[1,8,-1],[-1,1,-1]])
        # im_CustomFilter, lab_CustomFilter = CustomFilter(lpls)(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im_CustomFilter, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab_CustomFilter, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_Sharpen.png')
        # save_lab_path = os.path.join(save_label, name + '_Sharpen.png')
        # im__Sharpen, lab__Sharpen = Sharpen()(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im__Sharpen, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab__Sharpen, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_GaussianBlur.png')
        # save_lab_path = os.path.join(save_label, name + '_GaussianBlur.png')
        # im__GaussianBlur, lab__GaussianBlur = GaussianBlur()(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im__GaussianBlur, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab__GaussianBlur, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_Perspective.png')
        # save_lab_path = os.path.join(save_label, name + '_Perspective.png')
        # im__Perspective, lab__Perspective = Perspective()(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im__Perspective, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab__Perspective, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_Cutout.png')
        # save_lab_path = os.path.join(save_label, name + '_Cutout.png')
        # im__Cutout, lab__Cutout = Cutout([0.01,0.05],[0.2,0.3])(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im__Cutout, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab__Cutout, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_Leaf.png')
        # save_lab_path = os.path.join(save_label, name + '_Leaf.png')
        # im__Leaf, lab__Leaf = Leaf()(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im__Leaf, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab__Leaf, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_Brightness.png')
        # save_lab_path = os.path.join(save_label, name + '_Brightness.png')
        # im__Brightness, lab__Brightness = Brightness()(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im__Brightness, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab__Brightness, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_Contrast.png')
        # save_lab_path = os.path.join(save_label, name + '_Contrast.png')
        # im__Contrast, lab__Contrast = Contrast()(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im__Contrast, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab__Contrast, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_UniformNoise.png')
        # save_lab_path = os.path.join(save_label, name + '_UniformNoise.png')
        # im__UniformNoise, lab__UniformNoise = UniformNoise()(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im__UniformNoise, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab__UniformNoise, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_GaussianNoise.png')
        # save_lab_path = os.path.join(save_label, name + '_GaussianNoise.png')
        # im__GaussianNoise, lab__GaussianNoise = GaussianNoise()(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im__GaussianNoise, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab__GaussianNoise, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        # im_arr = cv2.imread(im_full_path)
        # lab_arr = cv2.imread(lab_full_path,0)
        # save_img_path = os.path.join(save_img, name + '_Vignetting.png')
        # save_lab_path = os.path.join(save_label, name + '_Vignetting.png')
        # im__Vignetting, lab__Vignetting = Vignetting()(im_arr, lab_arr)
        # cv2.imwrite(save_img_path, im__Vignetting, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        # cv2.imwrite(save_lab_path, lab__Vignetting, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

        im_arr = cv2.imread(im_full_path)
        lab_arr = cv2.imread(lab_full_path,0)
        save_img_path = os.path.join(save_img, name + '_LensDistortion.png')
        save_lab_path = os.path.join(save_label, name + '_LensDistortion.png')
        im__LensDistortion, lab__LensDistortion = LensDistortion()(im_arr, lab_arr)
        cv2.imwrite(save_img_path, im__LensDistortion, [int(cv2.IMWRITE_PNG_COMPRESSION),0])
        cv2.imwrite(save_lab_path, lab__LensDistortion, [int(cv2.IMWRITE_PNG_COMPRESSION),0])

if __name__ == "__main__":
    img_path = './crop/1/'
    label_path = './crop/1/'

    save_img = './crop/1/'
    if not os.path.exists(save_img):
        os.makedirs(save_img)
    save_label = './crop/1/'
    if not os.path.exists(save_label):
        os.makedirs(save_label)

    data_aug(img_path, label_path, save_img, save_label)

3.2 结果

对了,建议大家用彩色图片,效果明显,我直接用我论文的图了。
数据增强之后效果变差,激光条纹中心提取,python,python,计算机视觉,数据增强

a)Image;b)Resize;c)Clip;d)Normalize;f)Standardize;g)Flip;h)Crop i)Custom Filter;j)Sharpen;k)Gaussian Blur;l)Perspective;m)Cutout;n)Brightness;o)Contrast;p)Uniform Noise;q)Gaussian Noise

代码2文章来源地址https://www.toymoban.com/news/detail-719510.html

import torchvision.transforms as transforms

from PIL import Image

img = Image.open('tina.jpg')

'''
# CenterCrop
size = (224, 224)
transform = transforms.CenterCrop(size)
center_crop = transform(img)
center_crop.save('center_crop.jpg')
# ColorJitter
brightness = (1, 10)
contrast = (1, 10)
saturation = (1, 10)
hue = (0.2, 0.4)
transform = transforms.ColorJitter(brightness, contrast, saturation, hue)
color_jitter = transform(img)
color_jitter.save('color_jitter.jpg')
# FiveCrop
size = (224, 224)
transform = transforms.FiveCrop(size)
five_crop = transform(img)
for index, img in enumerate(five_crop):
    img.save(str(index) + '.jpg')
# Grayscale
transform = transforms.Grayscale()
grayscale = transform(img)
grayscale.save('grayscale.jpg')
# Compose, Pad
size = (224, 224)
padding = 16
fill = (0, 0, 255)
transform = transforms.Compose([
        transforms.CenterCrop(size),
        transforms.Pad(padding, fill)
])
pad = transform(img)
pad.save('pad.jpg')
# RandomAffine
degrees = (15, 30)
translate=(0, 0.2)
scale=(0.8, 1)
fillcolor = (0, 0, 255)
transform = transforms.RandomAffine(degrees=degrees, translate=translate, scale=scale, fillcolor=fillcolor)
random_affine = transform(img)
random_affine.save('random_affine.jpg')
# RandomApply
size = (224, 224)
padding = 16
fill = (0, 0, 255)
transform = transforms.RandomApply([transforms.CenterCrop(size), transforms.Pad(padding, fill)])
for i in range(3):
    random_apply = transform(img)
    random_apply.save(str(i) + '.jpg')
# RandomChoice
transform = transforms.RandomChoice([transforms.RandomAffine(degrees), 
                                     transforms.CenterCrop(size), 
                                     transforms.Pad(padding, fill)])
for i in range(3):
    random_order = transform(img)
    random_order.save(str(i) + '.jpg')
# RandomCrop
size = (224, 224)
transform = transforms.RandomCrop(size)
random_crop = transform(img)
random_crop.save('p.jpg')
# RandomGrayscale
p = 0.5
transform = transforms.RandomGrayscale(p)
for i in range(3):
    random_grayscale = transform(img)
    random_grayscale.save(str(i) + '.jpg')
# RandomHorizontalFlip
p = 0.5
transform = transforms.RandomHorizontalFlip(p)
for i in range(3):
    random_horizontal_filp = transform(img)
    random_horizontal_filp.save(str(i) + '.jpg')
# RandomOrder
size = (224, 224)
padding = 16
fill = (0, 0, 255)
degrees = (15, 30)
transform = transforms.RandomOrder([transforms.RandomAffine(degrees), 
                                    transforms.CenterCrop(size), 
                                    transforms.Pad(padding, fill)])
for i in range(3):
    random_order = transform(img)
    random_order.save(str(i) + '.jpg')
# RandomPerspective
distortion_scale = 1
p = 1
fill = (0, 0, 255)
transform = transforms.RandomPerspective(distortion_scale=distortion_scale, p=p, fill=fill)
random_perspective = transform(img)
random_perspective.save('random_perspective.jpg')
# RandomResizedCrop
size = (256, 256)
scale=(0.8, 1.0)
ratio=(0.75, 1.0)
transform = transforms.RandomResizedCrop(size=size, scale=scale, ratio=ratio)
random_resized_crop = transform(img)
random_resized_crop.save('random_resized_crop.jpg')
# RandomRotation
degrees = (15, 30)
fill = (0, 0, 255)
transform = transforms.RandomRotation(degrees=degrees, fill=fill)
random_rotation = transform(img)
random_rotation.save('random_rotation.jpg')
# RandomVerticalFlip
p = 1
transform = transforms.RandomVerticalFlip(p)
random_vertical_filp = transform(img)
random_vertical_filp.save('random_vertical_filp.jpg')
# Resize
size = (224, 224)
transform = transforms.Resize(size)
resize_img = transform(img)
resize_img.save('resize_img.jpg')
# ToPILImage
img = Image.open('tina.jpg')
transform = transforms.ToTensor()
img = transform(img)
print(img.size())
img_r = img[0, :, :]
img_g = img[1, :, :]
img_b = img[2, :, :]
print(type(img_r))
print(img_r.size())
transform = transforms.ToPILImage()
img_r = transform(img_r)
img_g = transform(img_g)
img_b = transform(img_b)
print(type(img_r))
img_r.save('img_r.jpg')
img_g.save('img_g.jpg')
img_b.save('img_b.jpg')
# ToTensor
img = Image.open('tina.jpg')
print(type(img))
print(img.size)
transform = transforms.ToTensor()
img = transform(img)
print(type(img))
print(img.size())
'''

# RandomErasing
p = 1.0
scale = (0.2, 0.3)
ratio = (0.5, 1.0)
value = (0, 0, 255)

transform = transforms.Compose([
                transforms.ToTensor(),
                transforms.RandomErasing(p=p, scale=scale, ratio=ratio, value=value),
                transforms.ToPILImage()
            ])
random_erasing = transform(img)
random_erasing.save('random_erasing.jpg')

到了这里,关于数据增强(Data Augmentation)的文章就介绍完了。如果您还想了解更多内容,请在右上角搜索TOY模板网以前的文章或继续浏览下面的相关文章,希望大家以后多多支持TOY模板网!

本文来自互联网用户投稿,该文观点仅代表作者本人,不代表本站立场。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如若转载,请注明出处: 如若内容造成侵权/违法违规/事实不符,请点击违法举报进行投诉反馈,一经查实,立即删除!

领支付宝红包 赞助服务器费用

相关文章

  • Dr. LLaMA: Improving Small Language Models in Domain-Specific QAvia Generative Data Augmentation

    https://arxiv.org/pdf/2305.07804.pdf https://arxiv.org/pdf/2305.07804.pdf Our findings indicate that LLMs effectively refine and diversify existing question-answer pairs, resulting in improved performance of a much smaller model on domain-specific QA datasets after fine-tuning.This study highlights the challenges of using LLMs for domain-specific question an

    2024年02月15日
    浏览(44)
  • RGB-L:基于激光雷达增强的ORB_SLAM3(已开源)

    点云PCL免费知识星球,点云论文速读。 文章:RGB-L: Enhancing Indirect Visual SLAM using LiDAR-based Dense Depth Maps 作者:Florian Sauerbeck, Benjamin Obermeier, Martin Rudolph 编辑:点云PCL 代码:https://github.com/TUMFTM/ORB_SLAM3_RGBL.git 欢迎各位加入免费知识星球,获取PDF论文,欢迎转发朋友圈。文章仅

    2024年02月07日
    浏览(48)
  • rime中州韵 输入效果一览 100+增强功能效果

    rime是一个定制化程度很高的输入法框架, 我们可以在该框架上搭建适合自己的输入法程序。我们将在专栏 小狼毫 Rime 保姆教程 中完成以下近百种定制化效果的配置与演示。欢迎订阅。 以下为个性化定制的输入效果: 👇 中文(五笔)输入, 英文(easy-english)输入, latex输入 单词首

    2024年02月03日
    浏览(43)
  • 使用增强版 singleflight 合并事件推送,效果炸裂!

    hello,大家好啊,我是小楼。 最近在工作中对 Go 的 singleflight 包做了下增强,解决了一个性能问题,这里记录下,希望对你也有所帮助。 singleflight 直接翻译为”单(次)飞(行)“,它是对同一种请求的抑制,保证同一时刻相同的请求只有一个在执行,且在它执行期间的相

    2024年02月05日
    浏览(32)
  • imgaug库图像增强指南(32):塑造【雪景】效果的视觉魔法

    在深度学习和计算机视觉的世界里, 数据是模型训练的基石 ,其质量与数量直接影响着模型的性能。然而,获取大量高质量的标注数据往往需要耗费大量的时间和资源。正因如此,数据增强技术应运而生,成为了解决这一问题的关键所在。而 imgaug ,作为 一个功能强大的图

    2024年01月25日
    浏览(41)
  • 为什么视频画质会变差,如何提升视频画质清晰度。

    在数字时代,视频已经成为我们生活中不可或缺的一部分。然而,随着视频的传输和处理过程中的多次压缩,画质损失逐渐凸显,影响了我们对影像的真实感受。为了让视频画质更加清晰、逼真,我们需要采取一些措施来保护和修复视频画质,还原影像的真实之美。 1、录制

    2024年02月16日
    浏览(42)
  • 自制易拉罐WiFi信号增强器详细图文教程(效果大揭秘)

    昨天iPhone中文网发布了一篇“简单六步增强WiFi信号 让你的iPhone如鱼得水”的教程,引来了不少果粉的质疑。针对“易拉罐是否能增强WiFi信号”这个饱受争议的命题,评测员决定亲手尝试,一来为大家揭开事实的真相,二来教大家快速制作易拉罐WiFi增强器,真正让你的iPhon

    2024年02月06日
    浏览(38)
  • uniapp 下拉框效果使用 uni-data-select标签

    uni-data-select v-model=\\\"formData.femMerchantsClassificationId\\\" :localdata=\\\"range\\\" @change=\\\"change\\\" /uni-data-select :localdata 绑定下拉框内容 当下拉框内容调用后端接口时候,写法:

    2024年02月11日
    浏览(44)
  • 第四篇:3.3 无效流量(Invalid traffic) - IAB/MRC及《增强现实广告效果测量指南1.0》

    翻译计划 第一篇 概述—IAB与MRC及《增强现实广告效果测量指南》之目录、适用范围及术语 第二篇 广告效果测量定义和其他矩阵之- 3.1 广告印象(AD Impression) 第三篇 广告效果测量定义和其他矩阵之- 3.2 可见性 (Viewability) 第四篇 广告效果测量定义和其他矩阵之- 3.3 无效流

    2024年04月12日
    浏览(27)
  • 第五篇:3.4 用户归因和受众(User attribution and audience) - IAB/MRC及《增强现实广告效果测量指南1.0》

    翻译计划 第一篇 概述—IAB与MRC及《增强现实广告效果测量指南》之目录、适用范围及术语 第二篇 广告效果测量定义和其他矩阵之- 3.1 广告印象(AD Impression) 第三篇 广告效果测量定义和其他矩阵之- 3.2 可见性 (Viewability) 第四篇 广告效果测量定义和其他矩阵之- 3.3 无效流

    2024年04月17日
    浏览(32)

觉得文章有用就打赏一下文章作者

支付宝扫一扫打赏

博客赞助

微信扫一扫打赏

请作者喝杯咖啡吧~博客赞助

支付宝扫一扫领取红包,优惠每天领

二维码1

领取红包

二维码2

领红包