博客
关于我
数据增强--对于目标检测(边框、图像)
阅读量:559 次
发布时间:2019-03-09

本文共 6607 字,大约阅读时间需要 22 分钟。

import torchfrom PIL import Image, ImageFont, ImageDrawfrom functools import reduceimport scipy.io as sciofrom PIL import Imageimport cv2 as cvimport numpy as npimport randomimport imutilsfrom imgaug import augmenters as iaadef horisontal_flip(images, targets):    images = torch.flip(images, [-1])    targets[:, 2] = 1 - targets[:, 2]    return images, targetsdef compose(*funcs):    """Compose arbitrarily many functions, evaluated left to right.    Reference: https://mathieularose.com/function-composition-in-python/    """    # return lambda x: reduce(lambda v, f: f(v), funcs, x)    if funcs:        return reduce(lambda f, g: lambda *a, **kw: g(f(*a, **kw)), funcs)    else:        raise ValueError('Composition of empty sequence not supported.')def letterbox_image(image, size):    '''resize image with unchanged aspect ratio using padding'''    iw, ih = image.size    w, h = size    scale = min(w / iw, h / ih)    nw = int(iw * scale)    nh = int(ih * scale)    image = image.resize((nw, nh), Image.BICUBIC)    new_image = Image.new('RGB', size, (128, 128, 128))    new_image.paste(image, ((w - nw) // 2, (h - nh) // 2))    return new_imagedef rand(a=0, b=1):    return np.random.rand() * (b - a) + adef resize_3D_data(ima, size=(416, 416)):    dat = np.zeros([416, 416, 15])    for i in range(15):        slice = np.squeeze(ima[:, :, i])        re_slice = cv.resize(slice.astype('uint8'), size, interpolation=cv.INTER_AREA)        dat[:, :, i] = re_slice    return datdef random_crop(cell, boxes):    dx = random.randint(15, 20)    dy = random.randint(15, 20)    shape = cell.shape    nx = shape[0]    ny = shape[1]    boxes[:, [0, 2]] = boxes[:, [0, 2]] - dx    boxes[:, [1, 3]] = boxes[:, [1, 3]] - dy    new_cell = np.zeros_like(cell)    new_cell[0:int(nx - dx), 0:int(ny - dy)] = cell[dx:, dy:]    new_cell = Image.fromarray(new_cell.astype('uint8')).convert('RGB')    new_cell = cv.cvtColor(np.asarray(new_cell), cv.COLOR_RGB2BGR)    boxes = np.where(boxes < 0, 0, boxes)    return new_cell, boxesdef random_noise(cell, boxes):    image = Image.fromarray(cell.astype('uint8')).convert('RGB')    image = cv.cvtColor(np.asarray(image), cv.COLOR_RGB2BGR)    seq = iaa.Sequential(        [            iaa.AdditiveGaussianNoise(scale=0.05 * 255),            iaa.LinearContrast((0.75, 1.5)),            iaa.GaussianBlur(sigma=(0, 4.0)),            iaa.Dropout(p=(0, 0.2)),            iaa.CoarseDropout(0.02, size_percent=0.5)        ], random_order=True    )    images_aug = seq.augment_images([image])[0]    new_cell = Image.fromarray(cv.cvtColor(images_aug, cv.COLOR_BGR2RGB))    return new_cell, boxesdef gray_level_crop(cell, boxes):    max_val = random.randint(80, 255) / 255    new_cell = cell * max_val    new_cell = Image.fromarray(new_cell.astype('uint8')).convert('RGB')    new_cell = cv.cvtColor(np.asarray(new_cell), cv.COLOR_RGB2BGR)    return new_cell, boxesdef rotate_box(box, M, shape):    # print(box)    y1, x1, y2, x2 = box    p1 = np.array([x1, y1, 1]).reshape((3, 1))    p2 = np.array([x1, y2, 1]).reshape((3, 1))    p3 = np.array([x2, y2, 1]).reshape((3, 1))    p4 = np.array([x2, y1, 1]).reshape((3, 1))    p1 = np.matmul(M, p1)    p2 = np.matmul(M, p2)    p3 = np.matmul(M, p3)    p4 = np.matmul(M, p4)    x1 = np.min([p1[0, 0], p2[0, 0], p3[0, 0], p4[0, 0]])    x2 = np.max([p1[0, 0], p2[0, 0], p3[0, 0], p4[0, 0]])    y1 = np.min([p1[1, 0], p2[1, 0], p3[1, 0], p4[1, 0]])    y2 = np.max([p1[1, 0], p2[1, 0], p3[1, 0], p4[1, 0]])    if x1 < 0:        x1 = 0    if x1 > shape[1]:        x1 = shape[1] - 1    if x2 < 0:        x2 = 0    if x2 > shape[1]:        x2 = shape[1] - 1    if y1 < 0:        y1 = 0    if y1 > shape[0]:        y1 = shape[0] - 1    if y2 < 0:        y2 = 0    if y2 > shape[0]:        y2 = shape[0] - 1    box = [y1, x1, y2, x2]    # print(box)    # print('--------------')    return boxdef random_rotate(cell, boxes, angle=45):    (h, w) = cell.shape    (cX, cY) = (w // 2, h // 2)    new_cell = imutils.rotate_bound(cell.astype('uint8'), angle)    M = cv.getRotationMatrix2D((cX, cY), -angle, 1.0)    cos = np.abs(M[0, 0])    sin = np.abs(M[0, 1])    # compute the new bounding dimensions of the image    nW = int((h * sin) + (w * cos))    nH = int((h * cos) + (w * sin))    # adjust the rotation matrix to take into account translation    M[0, 2] += (nW / 2) - cX    M[1, 2] += (nH / 2) - cY    new_boxes = []    for i in range(len(boxes)):        new_boxes.append(rotate_box(boxes[i], M, new_cell.shape))    if len(new_boxes) > 0:        new_boxes = np.array(new_boxes)    new_cell = Image.fromarray(new_cell.astype('uint8')).convert('RGB')    new_cell = cv.cvtColor(np.asarray(new_cell), cv.COLOR_RGB2BGR)    return new_cell, new_boxesdef random_argument(data, cell, boxes):    # crop region ...    if rand() < .5:        cell, boxes = random_crop(cell, boxes)    # add noise ...    if rand() < .5:        cell, boxes = random_noise(cell, boxes)    # add noise ...    if rand() < .5:        cell, boxes = gray_level_crop(cell, boxes)    # rotate ...    if rand() < .5:       cell, boxes = random_rotate(cell, boxes, random.randint(0, 90))    return cell, boxesdef pad_to_square(img, pad_value):    h, w, c = img.shape    dim_diff = np.abs(h - w)    # (upper / left) padding and (lower / right) padding    pad1, pad2 = dim_diff // 2, dim_diff - dim_diff // 2    # Determine padding    pad = ((0, 0), (pad1, pad2), (0, 0)) if w <= h else ((pad1, pad2), (0, 0), (0, 0))    # Add padding    img = np.pad(img, pad, "constant", constant_values=pad_value)    return img, padif __name__ == '__main__':    dat = scio.loadmat(        '/home/xuxu/Data/CTC_Signal_Datasets/Processed/20200616_cell_signalmark_39frames/Green/653802_650142_0_675.mat')    cell = dat['signal'].astype('uint8')    boxe = dat['rects']    print('boxe', boxe)    # image = Image.fromarray(cell.astype('uint8')).convert('RGB')    # image = cv.cvtColor(np.asarray(image), cv.COLOR_RGB2BGR)    # for x1, y1, x2, y2 in boxes:    #     cv.rectangle(image, (y1, x1), (y2, x2), (255, 255, 255), thickness=1)    # cv.imwrite('1.jpg', image)    new_cell, new_boxes = random_rotate(cell, boxe)    # new_cell, new_boxes = gray_level_crop(cell, boxe)    # print('new_boxes', new_boxes)    # new_boxes = np.where(new_boxes < 0, 0, new_boxes)    # print('new_boxes', new_boxes)    # image = Image.fromarray(new_cell.astype('uint8')).convert('RGB')    # image = cv.cvtColor(np.asarray(image), cv.COLOR_RGB2BGR)    for x1, y1, x2, y2 in new_boxes:        cv.rectangle(new_cell, (int(y1), int(x1)), (int(y2), int(x2)), (255, 255, 255), thickness=1)    cv.imwrite('1.jpg', new_cell)

其中图像加噪使用的是imgaug库

转载地址:http://ewhsz.baihongyu.com/

你可能感兴趣的文章
MySQL-数据页的结构
查看>>
MySQL-架构篇
查看>>
MySQL-索引的分类(聚簇索引、二级索引、联合索引)
查看>>
Mysql-触发器及创建触发器失败原因
查看>>
MySQL-连接
查看>>
mysql-递归查询(二)
查看>>
MySQL5.1安装
查看>>
mysql5.5和5.6版本间的坑
查看>>
mysql5.5最简安装教程
查看>>
mysql5.6 TIME,DATETIME,TIMESTAMP
查看>>
mysql5.6.21重置数据库的root密码
查看>>
Mysql5.6主从复制-基于binlog
查看>>
MySQL5.6忘记root密码(win平台)
查看>>
MySQL5.6的Linux安装shell脚本之二进制安装(一)
查看>>
MySQL5.6的zip包安装教程
查看>>
mysql5.7 for windows_MySQL 5.7 for Windows 解压缩版配置安装
查看>>
Webpack 基本环境搭建
查看>>
mysql5.7 安装版 表不能输入汉字解决方案
查看>>
MySQL5.7.18主从复制搭建(一主一从)
查看>>
MySQL5.7.19-win64安装启动
查看>>