百度飞桨深度学习识虫数据处理

数据集处理

AI识虫数据集结构如下：

提供了2183张图片，其中训练集1693张，验证集245，测试集245张。
包含7种昆虫，分别是Boerner、Leconte、Linnaeus、acuminatus、armandi、coleoptera和linnaeus。
包含了图片和标注，请先将数据解压，并存放在insects目录下。
在新建项目时可以选择数据集，有多个数据集可以使用。如下图

解压文件

# 解压数据脚本，第一次运行时打开注释，将文件解压到work目录下
# !unzip -d /home/aistudio/work /home/aistudio/data/data19638/insects.zip

调用get_insect_names函数返回一个dict，描述了昆虫名称和数字类别之间的映射关系。下面的程序从annotations/xml目录下面读取所有文件标注信息。

import os
import numpy as np
import xml.etree.ElementTree as ET

def get_annotations(cname2cid, datadir):
    filenames = os.listdir(os.path.join(datadir, 'annotations', 'xmls'))
    records = []
    ct = 0
    for fname in filenames:
        fid = fname.split('.')[0]
        fpath = os.path.join(datadir, 'annotations', 'xmls', fname)
        img_file = os.path.join(datadir, 'images', fid + '.jpeg')
        tree = ET.parse(fpath)

        if tree.find('id') is None:
            im_id = np.array([ct])
        else:
            im_id = np.array([int(tree.find('id').text)])

        objs = tree.findall('object')
        im_w = float(tree.find('size').find('width').text)
        im_h = float(tree.find('size').find('height').text)
        gt_bbox = np.zeros((len(objs), 4), dtype=np.float32)
        gt_class = np.zeros((len(objs), ), dtype=np.int32)
        is_crowd = np.zeros((len(objs), ), dtype=np.int32)
        difficult = np.zeros((len(objs), ), dtype=np.int32)
        for i, obj in enumerate(objs):
            cname = obj.find('name').text
            gt_class[i] = cname2cid[cname]
            _difficult = int(obj.find('difficult').text)
            x1 = float(obj.find('bndbox').find('xmin').text)
            y1 = float(obj.find('bndbox').find('ymin').text)
            x2 = float(obj.find('bndbox').find('xmax').text)
            y2 = float(obj.find('bndbox').find('ymax').text)
            x1 = max(0, x1)
            y1 = max(0, y1)
            x2 = min(im_w - 1, x2)
            y2 = min(im_h - 1, y2)
            # 这里使用xywh格式来表示目标物体真实框
            gt_bbox[i] = [(x1+x2)/2.0 , (y1+y2)/2.0, x2-x1+1., y2-y1+1.]
            is_crowd[i] = 0
            difficult[i] = _difficult

        voc_rec = {
            'im_file': img_file,
            'im_id': im_id,
            'h': im_h,
            'w': im_w,
            'is_crowd': is_crowd,
            'gt_class': gt_class,
            'gt_bbox': gt_bbox,
            'gt_poly': [],
            'difficult': difficult
            }
        if len(objs) != 0:
            records.append(voc_rec)
        ct += 1
    return records

数据读取

get_img_data_from_file()函数可以返回图片数据的数据，它们是图像数据img，真实框坐标gt_boxes，真实框包含的物体类别gt_labels，图像尺寸scales。

### 数据读取
import cv2

def get_bbox(gt_bbox, gt_class):
   # 对于一般的检测任务来说，一张图片上往往会有多个目标物体
   # 设置参数MAX_NUM = 50， 即一张图片最多取50个真实框；如果真实
   # 框的数目少于50个，则将不足部分的gt_bbox, gt_class和gt_score的各项数值全设置为0
   MAX_NUM = 50
   gt_bbox2 = np.zeros((MAX_NUM, 4))
   gt_class2 = np.zeros((MAX_NUM,))
   for i in range(len(gt_bbox)):
       gt_bbox2[i, :] = gt_bbox[i, :]
       gt_class2[i] = gt_class[i]
       if i >= MAX_NUM:
           break
   return gt_bbox2, gt_class2

def get_img_data_from_file(record):
   """
   record is a dict as following,
     record = {
           'im_file': img_file,
           'im_id': im_id,
           'h': im_h,
           'w': im_w,
           'is_crowd': is_crowd,
           'gt_class': gt_class,
           'gt_bbox': gt_bbox,
           'gt_poly': [],
           'difficult': difficult
           }
   """
   im_file = record['im_file']
   h = record['h']
   w = record['w']
   is_crowd = record['is_crowd']
   gt_class = record['gt_class']
   gt_bbox = record['gt_bbox']
   difficult = record['difficult']

   img = cv2.imread(im_file)
   img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

   # check if h and w in record equals that read from img
   assert img.shape[0] == int(h), \
            "image height of {} inconsistent in record({}) and img file({})".format(
              im_file, h, img.shape[0])

   assert img.shape[1] == int(w), \
            "image width of {} inconsistent in record({}) and img file({})".format(
              im_file, w, img.shape[1])

   gt_boxes, gt_labels = get_bbox(gt_bbox, gt_class)

   # gt_bbox 用相对值
   gt_boxes[:, 0] = gt_boxes[:, 0] / float(w)
   gt_boxes[:, 1] = gt_boxes[:, 1] / float(h)
   gt_boxes[:, 2] = gt_boxes[:, 2] / float(w)
   gt_boxes[:, 3] = gt_boxes[:, 3] / float(h)
 
   return img, gt_boxes, gt_labels, (h, w)

数据增强

什么是数据增强呢？data augmentation，它的意思是让有限的数据产生更多的等价数据。
由于数据集已经提供，我们只能从数据增强方面多添加训练数据。

随机改变亮暗、对比度和颜色等
图像亮度、对比度、饱和度和锐化之间并不是彼此独立的，改变其中一个特征可能会同时引起图像其他特征的变化，至于变化的程度取决于图像本身的特性。

图像亮度通俗理解便是图像的明暗程度，数字图像 f(x，y) = i(x，y) r(x， y) ，如果灰度值在[0，255]之间，则 f 值越接近0亮度越低，f 值越接近255亮度越高。

饱和度指的是图像颜色种类的多少， 上面提到图像的灰度级是[Lmin，Lmax]，则在Lmin、Lmax 的中间值越多，便代表图像的颜色种类多，饱和度也就更高，外观上看起来图像会更鲜艳，调整饱和度可以修正过度曝光或者未充分曝光的图片。使图像看上去更加自然。

对比度指的是图像暗和亮的落差值，即图像最大灰度级和最小灰度级之间的差值。

import numpy as np
import cv2
from PIL import Image, ImageEnhance
import random

# 随机改变亮暗、对比度和颜色等
def random_distort(img):
    # 随机改变亮度
    def random_brightness(img, lower=0.5, upper=1.5):
        e = np.random.uniform(lower, upper)
        return ImageEnhance.Brightness(img).enhance(e)
    # 随机改变对比度
    def random_contrast(img, lower=0.5, upper=1.5):
        e = np.random.uniform(lower, upper)
        return ImageEnhance.Contrast(img).enhance(e)
    # 随机改变颜色
    def random_color(img, lower=0.5, upper=1.5):
        e = np.random.uniform(lower, upper)
        return ImageEnhance.Color(img).enhance(e)

    ops = [random_brightness, random_contrast, random_color]
    np.random.shuffle(ops)

    img = Image.fromarray(img)
    img = ops[0](img)
    img = ops[1](img)
    img = ops[2](img)
    img = np.asarray(img)

    return img

随机填充

# 随机填充
def random_expand(img,
                  gtboxes,
                  max_ratio=4.,
                  fill=None,
                  keep_ratio=True,
                  thresh=0.5):
    if random.random() > thresh:
        return img, gtboxes

    if max_ratio < 1.0:
        return img, gtboxes

    h, w, c = img.shape
    ratio_x = random.uniform(1, max_ratio)
    if keep_ratio:
        ratio_y = ratio_x
    else:
        ratio_y = random.uniform(1, max_ratio)
    oh = int(h * ratio_y)
    ow = int(w * ratio_x)
    off_x = random.randint(0, ow - w)
    off_y = random.randint(0, oh - h)

    out_img = np.zeros((oh, ow, c))
    if fill and len(fill) == c:
        for i in range(c):
            out_img[:, :, i] = fill[i] * 255.0

    out_img[off_y:off_y + h, off_x:off_x + w, :] = img
    gtboxes[:, 0] = ((gtboxes[:, 0] * w) + off_x) / float(ow)
    gtboxes[:, 1] = ((gtboxes[:, 1] * h) + off_y) / float(oh)
    gtboxes[:, 2] = gtboxes[:, 2] / ratio_x
    gtboxes[:, 3] = gtboxes[:, 3] / ratio_y

    return out_img.astype('uint8'), gtboxes

随机裁剪
翻转操作和旋转操作，对于那些对方向不敏感的任务，比如图像分类，都是很常见的操作，在caffe等框架中翻转对应的就是mirror操作。翻转和旋转不改变图像的大小，而裁剪会改变图像的大小。通常在训练的时候会采用随机裁剪的方法，在测试的时候选择裁剪中间部分或者不裁剪。
值得注意的是，在一些竞赛中进行模型测试时，一般都是裁剪输入的多个版本然后将结果进行融合，对预测的改进效果非常明显。以上操作都不会产生失真，而缩放变形则是失真的。很多的时候，网络的训练输入大小是固定的，但是数据集中的图像却大小不一，此时就可以选择上面的裁剪成固定大小输入或者缩放到网络的输入大小的方案，后者就会产生失真，通常效果比前者差。

随机裁剪之前需要先定义两个函数，multi_box_iou_xywh和box_crop这两个函数将被保存在box_utils.py文件中。

import numpy as np

def multi_box_iou_xywh(box1, box2):
    """
    In this case, box1 or box2 can contain multi boxes.
    Only two cases can be processed in this method:
       1, box1 and box2 have the same shape, box1.shape == box2.shape
       2, either box1 or box2 contains only one box, len(box1) == 1 or len(box2) == 1
    If the shape of box1 and box2 does not match, and both of them contain multi boxes, it will be wrong.
    """
    assert box1.shape[-1] == 4, "Box1 shape[-1] should be 4."
    assert box2.shape[-1] == 4, "Box2 shape[-1] should be 4."


    b1_x1, b1_x2 = box1[:, 0] - box1[:, 2] / 2, box1[:, 0] + box1[:, 2] / 2
    b1_y1, b1_y2 = box1[:, 1] - box1[:, 3] / 2, box1[:, 1] + box1[:, 3] / 2
    b2_x1, b2_x2 = box2[:, 0] - box2[:, 2] / 2, box2[:, 0] + box2[:, 2] / 2
    b2_y1, b2_y2 = box2[:, 1] - box2[:, 3] / 2, box2[:, 1] + box2[:, 3] / 2

    inter_x1 = np.maximum(b1_x1, b2_x1)
    inter_x2 = np.minimum(b1_x2, b2_x2)
    inter_y1 = np.maximum(b1_y1, b2_y1)
    inter_y2 = np.minimum(b1_y2, b2_y2)
    inter_w = inter_x2 - inter_x1
    inter_h = inter_y2 - inter_y1
    inter_w = np.clip(inter_w, a_min=0., a_max=None)
    inter_h = np.clip(inter_h, a_min=0., a_max=None)

    inter_area = inter_w * inter_h
    b1_area = (b1_x2 - b1_x1) * (b1_y2 - b1_y1)
    b2_area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1)

    return inter_area / (b1_area + b2_area - inter_area)

def box_crop(boxes, labels, crop, img_shape):
    x, y, w, h = map(float, crop)
    im_w, im_h = map(float, img_shape)

    boxes = boxes.copy()
    boxes[:, 0], boxes[:, 2] = (boxes[:, 0] - boxes[:, 2] / 2) * im_w, (
        boxes[:, 0] + boxes[:, 2] / 2) * im_w
    boxes[:, 1], boxes[:, 3] = (boxes[:, 1] - boxes[:, 3] / 2) * im_h, (
        boxes[:, 1] + boxes[:, 3] / 2) * im_h

    crop_box = np.array([x, y, x + w, y + h])
    centers = (boxes[:, :2] + boxes[:, 2:]) / 2.0
    mask = np.logical_and(crop_box[:2] <= centers, centers <= crop_box[2:]).all(
        axis=1)

    boxes[:, :2] = np.maximum(boxes[:, :2], crop_box[:2])
    boxes[:, 2:] = np.minimum(boxes[:, 2:], crop_box[2:])
    boxes[:, :2] -= crop_box[:2]
    boxes[:, 2:] -= crop_box[:2]

    mask = np.logical_and(mask, (boxes[:, :2] < boxes[:, 2:]).all(axis=1))
    boxes = boxes * np.expand_dims(mask.astype('float32'), axis=1)
    labels = labels * mask.astype('float32')
    boxes[:, 0], boxes[:, 2] = (boxes[:, 0] + boxes[:, 2]) / 2 / w, (
        boxes[:, 2] - boxes[:, 0]) / w
    boxes[:, 1], boxes[:, 3] = (boxes[:, 1] + boxes[:, 3]) / 2 / h, (
        boxes[:, 3] - boxes[:, 1]) / h

    return boxes, labels, mask.sum()

# 随机裁剪
def random_crop(img,
                boxes,
                labels,
                scales=[0.3, 1.0],
                max_ratio=2.0,
                constraints=None,
                max_trial=50):
    if len(boxes) == 0:
        return img, boxes

    if not constraints:
        constraints = [(0.1, 1.0), (0.3, 1.0), (0.5, 1.0), (0.7, 1.0),
                       (0.9, 1.0), (0.0, 1.0)]

    img = Image.fromarray(img)
    w, h = img.size
    crops = [(0, 0, w, h)]
    for min_iou, max_iou in constraints:
        for _ in range(max_trial):
            scale = random.uniform(scales[0], scales[1])
            aspect_ratio = random.uniform(max(1 / max_ratio, scale * scale), \
                                          min(max_ratio, 1 / scale / scale))
            crop_h = int(h * scale / np.sqrt(aspect_ratio))
            crop_w = int(w * scale * np.sqrt(aspect_ratio))
            crop_x = random.randrange(w - crop_w)
            crop_y = random.randrange(h - crop_h)
            crop_box = np.array([[(crop_x + crop_w / 2.0) / w,
                                  (crop_y + crop_h / 2.0) / h,
                                  crop_w / float(w), crop_h / float(h)]])

            iou = multi_box_iou_xywh(crop_box, boxes)
            if min_iou <= iou.min() and max_iou >= iou.max():
                crops.append((crop_x, crop_y, crop_w, crop_h))
                break

    while crops:
        crop = crops.pop(np.random.randint(0, len(crops)))
        crop_boxes, crop_labels, box_num = box_crop(boxes, labels, crop, (w, h))
        if box_num < 1:
            continue
        img = img.crop((crop[0], crop[1], crop[0] + crop[2],
                        crop[1] + crop[3])).resize(img.size, Image.LANCZOS)
        img = np.asarray(img)
        return img, crop_boxes, crop_labels
    img = np.asarray(img)
    return img, boxes, labels

随机缩放

# 随机缩放
def random_interp(img, size, interp=None):
    interp_method = [
        cv2.INTER_NEAREST,
        cv2.INTER_LINEAR,
        cv2.INTER_AREA,
        cv2.INTER_CUBIC,
        cv2.INTER_LANCZOS4,
    ]
    if not interp or interp not in interp_method:
        interp = interp_method[random.randint(0, len(interp_method) - 1)]
    h, w, _ = img.shape
    im_scale_x = size / float(w)
    im_scale_y = size / float(h)
    img = cv2.resize(
        img, None, None, fx=im_scale_x, fy=im_scale_y, interpolation=interp)
    return img

随机翻转

# 随机翻转
def random_flip(img, gtboxes, thresh=0.5):
    if random.random() > thresh:
        img = img[:, ::-1, :]
        gtboxes[:, 0] = 1.0 - gtboxes[:, 0]
    return img, gtboxes

随机打乱真实框排列顺序

# 随机打乱真实框排列顺序
def shuffle_gtbox(gtbox, gtlabel):
    gt = np.concatenate(
        [gtbox, gtlabel[:, np.newaxis]], axis=1)
    idx = np.arange(gt.shape[0])
    np.random.shuffle(idx)
    gt = gt[idx, :]
    return gt[:, :4], gt[:, 4]

图像增广方法汇总

# 图像增广方法汇总
def image_augment(img, gtboxes, gtlabels, size, means=None):
    # 随机改变亮暗、对比度和颜色等
    img = random_distort(img)
    # 随机填充
    img, gtboxes = random_expand(img, gtboxes, fill=means)
    # 随机裁剪
    img, gtboxes, gtlabels, = random_crop(img, gtboxes, gtlabels)
    # 随机缩放
    img = random_interp(img, size)
    # 随机翻转
    img, gtboxes = random_flip(img, gtboxes)
    # 随机打乱真实框排列顺序
    gtboxes, gtlabels = shuffle_gtbox(gtboxes, gtlabels)

    return img.astype('float32'), gtboxes.astype('float32'), gtlabels.astype('int32')

将上面的过程整理成一个get_img_data函数。

def get_img_data(record, size=640):
    img, gt_boxes, gt_labels, scales = get_img_data_from_file(record)
    img, gt_boxes, gt_labels = image_augment(img, gt_boxes, gt_labels, size)
    mean = [0.485, 0.456, 0.406]
    std = [0.229, 0.224, 0.225]
    mean = np.array(mean).reshape((1, 1, -1))
    std = np.array(std).reshape((1, 1, -1))
    img = (img / 255.0 - mean) / std
    img = img.astype('float32').transpose((2, 0, 1))
    return img, gt_boxes, gt_labels, scales

TRAINDIR = '/home/aistudio/work/insects/train'
TESTDIR = '/home/aistudio/work/insects/test'
VALIDDIR = '/home/aistudio/work/insects/val'
cname2cid = get_insect_names()
records = get_annotations(cname2cid, TRAINDIR)

record = records[0]
img, gt_boxes, gt_labels, scales = get_img_data(record, size=480)

# 获取一个批次内样本随机缩放的尺寸
def get_img_size(mode):
    if (mode == 'train') or (mode == 'valid'):
        inds = np.array([0,1,2,3,4,5,6,7,8,9])
        ii = np.random.choice(inds)
        img_size = 320 + ii * 32
    else:
        img_size = 608
    return img_size

# 将 list形式的batch数据 转化成多个array构成的tuple
def make_array(batch_data):
    img_array = np.array([item[0] for item in batch_data], dtype = 'float32')
    gt_box_array = np.array([item[1] for item in batch_data], dtype = 'float32')
    gt_labels_array = np.array([item[2] for item in batch_data], dtype = 'int32')
    img_scale = np.array([item[3] for item in batch_data], dtype='int32')
    return img_array, gt_box_array, gt_labels_array, img_scale

# 批量读取数据，同一批次内图像的尺寸大小必须是一样的，
# 不同批次之间的大小是随机的，
# 由上面定义的get_img_size函数产生
def data_loader(datadir, batch_size= 10, mode='train'):
    cname2cid = get_insect_names()
    records = get_annotations(cname2cid, datadir)

    def reader():
        if mode == 'train':
            np.random.shuffle(records)
        batch_data = []
        img_size = get_img_size(mode)
        for record in records:
            #print(record)
            img, gt_bbox, gt_labels, im_shape = get_img_data(record, 
                                                             size=img_size)
            batch_data.append((img, gt_bbox, gt_labels, im_shape))
            if len(batch_data) == batch_size:
                yield make_array(batch_data)
                batch_data = []
                img_size = get_img_size(mode)
        if len(batch_data) > 0:
            yield make_array(batch_data)

    return reader

由于数据预处理耗时较长，可能会成为网络训练速度的瓶颈，所以需要对预处理部分进行优化。通过使用飞桨提供的API paddle.reader.xmap_readers可以开启多线程读取数据，具体实现代码如下。

import functools
import paddle

# 使用paddle.reader.xmap_readers实现多线程读取数据
def multithread_loader(datadir, batch_size= 10, mode='train'):
    cname2cid = get_insect_names()
    records = get_annotations(cname2cid, datadir)
    def reader():
        if mode == 'train':
            np.random.shuffle(records)
        img_size = get_img_size(mode)
        batch_data = []
        for record in records:
            batch_data.append((record, img_size))
            if len(batch_data) == batch_size:
                yield batch_data
                batch_data = []
                img_size = get_img_size(mode)
        if len(batch_data) > 0:
            yield batch_data

    def get_data(samples):
        batch_data = []
        for sample in samples:
            record = sample[0]
            img_size = sample[1]
            img, gt_bbox, gt_labels, im_shape = get_img_data(record, size=img_size)
            batch_data.append((img, gt_bbox, gt_labels, im_shape))
        return make_array(batch_data)

    mapper = functools.partial(get_data, )

    return paddle.reader.xmap_readers(mapper, reader, 8, 10)

至此，我们完成了如何查看数据集中的数据、提取数据标注信息、从文件读取图像和标注数据、图像增广、批量读取和加速等过程，通过multithread_loader可以返回img, gt_boxes, gt_labels, im_shape等数据，接下来就可以将它们输入到神经网络，应用到具体算法上了。

# 测试数据读取

# 将 list形式的batch数据 转化成多个array构成的tuple
def make_test_array(batch_data):
    img_name_array = np.array([item[0] for item in batch_data])
    img_data_array = np.array([item[1] for item in batch_data], dtype = 'float32')
    img_scale_array = np.array([item[2] for item in batch_data], dtype='int32')
    return img_name_array, img_data_array, img_scale_array

# 测试数据读取
def test_data_loader(datadir, batch_size= 10, test_image_size=608, mode='test'):
    """
    加载测试用的图片，测试数据没有groundtruth标签
    """
    image_names = os.listdir(datadir)
    def reader():
        batch_data = []
        img_size = test_image_size
        for image_name in image_names:
            file_path = os.path.join(datadir, image_name)
            img = cv2.imread(file_path)
            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
            H = img.shape[0]
            W = img.shape[1]
            img = cv2.resize(img, (img_size, img_size))

            mean = [0.485, 0.456, 0.406]
            std = [0.229, 0.224, 0.225]
            mean = np.array(mean).reshape((1, 1, -1))
            std = np.array(std).reshape((1, 1, -1))
            out_img = (img / 255.0 - mean) / std
            out_img = out_img.astype('float32').transpose((2, 0, 1))
            img = out_img #np.transpose(out_img, (2,0,1))
            im_shape = [H, W]

            batch_data.append((image_name.split('.')[0], img, im_shape))
            if len(batch_data) == batch_size:
                yield make_test_array(batch_data)
                batch_data = []
        if len(batch_data) > 0:
            yield make_test_array(batch_data)

    return reader

参考链接：

https://www.zhihu.com/question/319291048/answer/645047014