|
| 1 | +''' |
| 2 | + |
| 3 | +This is the part of https://github.com/githubharald/SimpleHTR with simple modification |
| 4 | +See License. |
| 5 | +''' |
| 6 | + |
| 7 | +from __future__ import division |
| 8 | +from __future__ import print_function |
| 9 | + |
| 10 | + |
| 11 | +import random |
| 12 | +import os |
| 13 | +import cv2 |
| 14 | +import numpy as np |
| 15 | + |
| 16 | +from SamplePreprocessor import preprocessor |
| 17 | + |
| 18 | + |
| 19 | +class FilePaths: |
| 20 | + """ Filenames and paths to data """ |
| 21 | + fnCharList = '../model/charList.txt' |
| 22 | + fnWordCharList = '../model/wordCharList.txt' |
| 23 | + fnCorpus = '../data/corpus.txt' |
| 24 | + fnAccuracy = '../model/accuracy.txt' |
| 25 | + fnTrain = '../data/' |
| 26 | + fnInfer = '../data/testImage1.png' ## path to recognize the single image |
| 27 | + |
| 28 | + |
| 29 | +class Sample: |
| 30 | + """ Sample from the dataset """ |
| 31 | + |
| 32 | + def __init__(self, gtText, filePath): |
| 33 | + self.gtText = gtText |
| 34 | + self.filePath = filePath |
| 35 | + |
| 36 | + |
| 37 | +class Batch: |
| 38 | + """ Batch containing images and ground truth texts """ |
| 39 | + |
| 40 | + def __init__(self, gtTexts, imgs): |
| 41 | + self.imgs = np.stack(imgs, axis=0) |
| 42 | + self.gtTexts = gtTexts |
| 43 | + |
| 44 | + |
| 45 | +class DataLoader: |
| 46 | + "loads data which corresponds to IAM format, see: http://www.fki.inf.unibe.ch/databases/iam-handwriting-database" |
| 47 | + |
| 48 | + def __init__(self, filePath, batchSize, imgSize, maxTextLen, load_aug=True): |
| 49 | + "loader for dataset at given location, preprocess images and text according to parameters" |
| 50 | + |
| 51 | + assert filePath[-1] == '/' |
| 52 | + |
| 53 | + self.dataAugmentation = True # False |
| 54 | + self.currIdx = 0 |
| 55 | + self.batchSize = batchSize |
| 56 | + self.imgSize = imgSize |
| 57 | + self.samples = [] |
| 58 | + |
| 59 | + f = open("../data/" + 'lines.txt') |
| 60 | + chars = set() |
| 61 | + bad_samples = [] |
| 62 | + bad_samples_reference = ['a01-117-05-02.png', 'r06-022-03-05.png'] |
| 63 | + for line in f: |
| 64 | + # ignore comment line |
| 65 | + if not line or line[0] == '#': |
| 66 | + continue |
| 67 | + |
| 68 | + lineSplit = line.strip().split(' ') ## remove the space and split with ' ' |
| 69 | + # assert len(lineSplit) >= 9 |
| 70 | + |
| 71 | + # filename: part1-part2-part3 --> part1/part1-part2/part1-part2-part3.png |
| 72 | + fileNameSplit = lineSplit[0].split('-') |
| 73 | + #print(fileNameSplit) |
| 74 | + fileName = filePath + 'lines/' + fileNameSplit[0] + '/' + fileNameSplit[0] + '-' + fileNameSplit[1] + '/' +\ |
| 75 | + lineSplit[0] + '.png' |
| 76 | + |
| 77 | + # GT text are columns starting at 10 |
| 78 | + # see the lines.txt and check where the GT text starts, in this case it is 10 |
| 79 | + gtText_list = lineSplit[9].split('|') |
| 80 | + gtText = self.truncateLabel(' '.join(gtText_list), maxTextLen) |
| 81 | + chars = chars.union(set(list(gtText))) ## taking the unique characters present |
| 82 | + |
| 83 | + # check if image is not empty |
| 84 | + if not os.path.getsize(fileName): |
| 85 | + bad_samples.append(lineSplit[0] + '.png') |
| 86 | + continue |
| 87 | + |
| 88 | + # put sample into list |
| 89 | + self.samples.append(Sample(gtText, fileName)) |
| 90 | + |
| 91 | + |
| 92 | + # some images in the IAM dataset are known to be damaged, don't show warning for them |
| 93 | + if set(bad_samples) != set(bad_samples_reference): |
| 94 | + print("Warning, damaged images found:", bad_samples) |
| 95 | + print("Damaged images expected:", bad_samples_reference) |
| 96 | + |
| 97 | + # split into training and validation set: 95% - 10% |
| 98 | + splitIdx = int(0.95 * len(self.samples)) |
| 99 | + self.trainSamples = self.samples[:splitIdx] |
| 100 | + self.validationSamples = self.samples[splitIdx:] |
| 101 | + print("Train: {}, Validation: {}".format(len(self.trainSamples), len(self.validationSamples))) |
| 102 | + # put lines into lists |
| 103 | + self.trainLines = [x.gtText for x in self.trainSamples] |
| 104 | + self.validationLines = [x.gtText for x in self.validationSamples] |
| 105 | + |
| 106 | + # number of randomly chosen samples per epoch for training |
| 107 | + self.numTrainSamplesPerEpoch = 9500 |
| 108 | + |
| 109 | + # start with train set |
| 110 | + self.trainSet() |
| 111 | + |
| 112 | + # list of all chars in dataset |
| 113 | + self.charList = sorted(list(chars)) |
| 114 | + |
| 115 | + def truncateLabel(self, text, maxTextLen): |
| 116 | + # ctc_loss can't compute loss if it cannot find a mapping between text label and input |
| 117 | + # labels. Repeat letters cost double because of the blank symbol needing to be inserted. |
| 118 | + # If a too-long label is provided, ctc_loss returns an infinite gradient |
| 119 | + cost = 0 |
| 120 | + for i in range(len(text)): |
| 121 | + if i != 0 and text[i] == text[i - 1]: |
| 122 | + cost += 2 |
| 123 | + else: |
| 124 | + cost += 1 |
| 125 | + if cost > maxTextLen: |
| 126 | + return text[:i] |
| 127 | + return text |
| 128 | + |
| 129 | + |
| 130 | + def trainSet(self): |
| 131 | + "switch to randomly chosen subset of training set" |
| 132 | + self.dataAugmentation = True |
| 133 | + self.currIdx = 0 |
| 134 | + random.shuffle(self.trainSamples) # shuffle the samples in each epoch |
| 135 | + self.samples = self.trainSamples #[:self.numTrainSamplesPerEpoch] |
| 136 | + |
| 137 | + def validationSet(self): |
| 138 | + "switch to validation set" |
| 139 | + self.dataAugmentation = False |
| 140 | + self.currIdx = 0 |
| 141 | + self.samples = self.validationSamples |
| 142 | + |
| 143 | + def getIteratorInfo(self): |
| 144 | + "current batch index and overall number of batches" |
| 145 | + return (self.currIdx // self.batchSize + 1, len(self.samples) // self.batchSize) |
| 146 | + |
| 147 | + def hasNext(self): |
| 148 | + "iterator" |
| 149 | + return self.currIdx + self.batchSize <= len(self.samples) |
| 150 | + |
| 151 | + def getNext(self): |
| 152 | + "iterator" |
| 153 | + batchRange = range(self.currIdx, self.currIdx + self.batchSize) |
| 154 | + gtTexts = [self.samples[i].gtText for i in batchRange] |
| 155 | + imgs = [preprocessor(cv2.imread(self.samples[i].filePath, cv2.IMREAD_GRAYSCALE), self.imgSize) |
| 156 | + for i in batchRange] |
| 157 | + self.currIdx += self.batchSize |
| 158 | + return Batch(gtTexts, imgs) |
0 commit comments