CRNN是OCR领域非常经典且被广泛使用的识别算法，其理论基础可以参考我上一篇文章，本文将着重讲解CRNN代码实现过程以及识别效果。
数据处理
利用图像处理技术我们手工大批量生成文字图像，一共360万张图像样本，效果如下：
我们划分了训练集和测试集（10:1），并单独存储为两个文本文件：
文本文件里的标签格式如下：
我们获取到的是最原始的数据集，在图像深度学习训练中我们一般都会把原始数据集转化为lmdb格式以方便后续的网络训练。因此我们也需要对该数据集进行lmdb格式转化。下面代码就是用于lmdb格式转化，思路比较简单，就是首先读入图像和对应的文本标签，先使用字典将该组合存储起来（cache），再利用lmdb包的put函数把字典(cache)存储的k,v写成lmdb格式存储好（cache当有了1000个元素就put一次）。
import lmdbimport cv2import numpy as npimport osdef checkImageIsValid(imageBin):if imageBin is None:return Falsetry:imageBuf = np.fromstring(imageBin, dtype=np.uint8)img = cv2.imdecode(imageBuf, cv2.IMREAD_GRAYSCALE)imgH, imgW = img.shape[0], img.shape[1]except:return Falseelse:if imgH * imgW == 0:return Falsereturn Truedef writeCache(env, cache):with env.begin(write=True) as txn:for k, v in cache.items():txn.put(k, v)def createDataset(outputPath, imagePathList, labelList, lexiconList=None, checkValid=True):"""Create LMDB dataset for CRNN training.ARGS:outputPath : LMDB output pathimagePathList : list of image pathlabelList: list of corresponding groundtruth textslexiconList : (optional) list of lexicon listscheckValid : if true, check the validity of every image"""assert (len(imagePathList) == len(labelList))nSamples = len(imagePathList)env = lmdb.open(outputPath, map_size=1099511627776)cache = {}cnt = 1for i in range(nSamples):imagePath = ''.join(imagePathList[i]).split()[0].replace('\n', '').replace('\r\n', '')# print(imagePath)label = ''.join(labelList[i])print(label)# if not os.path.exists(imagePath):#print('%s does not exist' % imagePath)#continuewith open('.' + imagePath, 'r') as f:imageBin = f.read()if checkValid:if not checkImageIsValid(imageBin):print('%s is not a valid image' % imagePath)continueimageKey = 'image-%09d' % cntlabelKey = 'label-%09d' % cntcache[imageKey] = imageBincache[labelKey] = labelif lexiconList:lexiconKey = 'lexicon-%09d' % cntcache[lexiconKey] = ' '.join(lexiconList[i])if cnt % 1000 == 0:writeCache(env, cache)cache = {}print('Written %d / %d' % (cnt, nSamples))cnt += 1print(cnt)nSamples = cnt - 1cache['num-samples'] = str(nSamples)writeCache(env, cache)print('Created dataset with %d samples' % nSamples)OUT_PATH = '../crnn_train_lmdb'IN_PATH = './train.txt'if __name__ == '__main__':outputPath = OUT_PATHif not os.path.exists(OUT_PATH):os.mkdir(OUT_PATH)imgdata = open(IN_PATH)imagePathList = list(imgdata)labelList = []for line in imagePathList:word = line.split()[1]labelList.append(word)createDataset(outputPath, imagePathList, labelList)
我们运行上面的代码，可以得到训练集和测试集的lmdb
在数据准备部分还有一个操作需要强调的，那就是文字标签数字化，即我们用数字来表示每一个文字（汉字，英文字母，标点符号）。比如“我”字对应的id是1，“l”对应的id是1000，“？”对应的id是90，如此类推，这种编解码工作使用字典数据结构存储即可，训练时先把标签编码（encode），预测时就将网络输出结果解码(decode)成文字输出。
class strLabelConverter(object):"""Convert between str and label.NOTE:Insert `blank` to the alphabet for CTC.Args:alphabet (str): set of the possible characters.ignore_case (bool, default=True): whether or not to ignore all of the case."""def __init__(self, alphabet, ignore_case=False):self._ignore_case = ignore_caseif self._ignore_case:alphabet = alphabet.lower()self.alphabet = alphabet + '-' # for `-1` indexself.dict = {}for i, char in enumerate(alphabet):# NOTE: 0 is reserved for 'blank' required by wrap_ctcself.dict[char] = i + 1def encode(self, text):"""Support batch or single str.Args:text (str or list of str): texts to convert.Returns:torch.IntTensor [length_0 + length_1 + ... length_{n - 1}]: encoded texts.torch.IntTensor [n]: length of each text."""length = []result = []for item in text:item = item.decode('utf-8', 'strict')length.append(len(item))for char in item:index = self.dict[char]result.append(index)text = result# print(text,length)return (torch.IntTensor(text), torch.IntTensor(length))def decode(self, t, length, raw=False):"""Decode encoded texts back into strs.Args:torch.IntTensor [length_0 + length_1 + ... length_{n - 1}]: encoded texts.torch.IntTensor [n]: length of each text.Raises:AssertionError: when the texts and its length does not match.Returns:text (str or list of str): texts to convert."""if length.numel() == 1:length = length[0]assert t.numel() == length, "text with length: {} does not match declared length: {}".format(t.numel(),length)if raw:return ''.join([self.alphabet[i - 1] for i in t])else:char_list = []for i in range(length):if t[i] != 0 and (not (i > 0 and t[i - 1] == t[i])):char_list.append(self.alphabet[t[i] - 1])return ''.join(char_list)else:# batch modeassert t.numel() == length.sum(), "texts with length: {} does not match declared length: {}".format(t.numel(), length.sum())texts = []index = 0for i in range(length.numel()):l = length[i]texts.append(self.decode(t[index:index + l], torch.IntTensor([l]), raw=raw))index += lreturn texts
网络设计
根据CRNN的论文描述，CRNN是由CNN-》RNN-》CTC三大部分架构而成，分别对应卷积层、循环层和转录层。首先CNN部分用于底层的特征提取，RNN采取了BiLSTM，用于学习关联序列信息并预测标签分布，CTC用于序列对齐，输出预测结果。
为了将特征输入到Recurrent Layers，做如下处理：
首先会将图像缩放到 32×W×3 大小然后经过CNN后变为 1×（W/4）× 512接着针对LSTM，设置 T=(W/4) ， D=512 ，即可将特征输入LSTM。
以上是理想训练时的操作，但是CRNN论文提到的网络输入是归一化好的100×32大小的灰度图像，即高度统一为32个像素。下面是CRNN的深度神经网络结构图，CNN采取了经典的VGG16，值得注意的是，在VGG16的第3第4个max pooling层CRNN采取的是1×2的矩形池化窗口(w×h)，这有别于经典的VGG16的2×2的正方形池化窗口，这个改动是因为文本图像多数都是高较小而宽较长，所以其feature map也是这种高小宽长的矩形形状，如果使用1×2的池化窗口则更适合英文字母识别（比如区分i和l）。VGG16部分还引入了BatchNormalization模块，旨在加速模型收敛。还有值得注意一点，CRNN的输入是灰度图像，即图像深度为1。CNN部分的输出是512x1x16（c×h×w）的特征向量。
接下来分析RNN层。RNN部分使用了双向LSTM，隐藏层单元数为256，CRNN采用了两层BiLSTM来组成这个RNN层，RNN层的输出维度将是（s,b,class_num） ，其中class_num为文字类别总数。
值得注意的是：Pytorch里的LSTM单元接受的输入都必须是3维的张量（Tensors）.每一维代表的意思不能弄错。第一维体现的是序列（sequence）结构，第二维度体现的是小块（mini-batch）结构，第三位体现的是输入的元素（elements of input）。如果在应用中不适用小块结构，那么可以将输入的张量中该维度设为1，但必须要体现出这个维度。
LSTM的输入
input of shape (seq_len, batch, input_size): tensor containing the features of the input sequence. The input can also be a packed variable length sequence.input shape(a,b,c)a:seq_len -> 序列长度b:batchc:input_size 输入特征数目
根据LSTM的输入要求，我们要对CNN的输出做些调整，即把CNN层的输出调整为[seq_len, batch, input_size]形式，下面为具体操作：先使用squeeze函数移除h维度，再使用permute函数调整各维顺序，即从原来[w, b, c]的调整为[seq_len, batch, input_size]，具体尺寸为[16,batch,512]，调整好之后即可以将该矩阵送入RNN层。
x = n(x)b, c, h, w = x.size()# print(x.size()): b,c,h,wassert h == 1 # "the height of conv must be 1"x = x.squeeze(2) # remove h dimension, b *512 * widthx = x.permute(2, 0, 1) # [w, b, c] = [seq_len, batch, input_size]x = self.rnn(x)
RNN层输出格式如下，因为我们采用的是双向BiLSTM，所以输出维度将是hidden_unit * 2
Outputs: output, (h_n, c_n)output of shape (seq_len, batch, num_directions * hidden_size)h_n of shape (num_layers * num_directions, batch, hidden_size)c_n (num_layers * num_directions, batch, hidden_size)
然后我们再通过线性变换操作self.embedding1 = torch.nn.Linear(hidden_unit * 2, 512)是的输出维度再次变为512，继续送入第二个LSTM层。第二个LSTM层后继续接线性操作torch.nn.Linear(hidden_unit * 2, class_num)使得整个RNN层的输出为文字类别总数。
import torchimport torch.nn.functional as Fclass Vgg_16(torch.nn.Module):def __init__(self):super(Vgg_16, self).__init__()self.convolution1 = torch.nn.Conv2d(1, 64, 3, padding=1)self.pooling1 = torch.nn.MaxPool2d(2, stride=2)self.convolution2 = torch.nn.Conv2d(64, 128, 3, padding=1)self.pooling2 = torch.nn.MaxPool2d(2, stride=2)self.convolution3 = torch.nn.Conv2d(128, 256, 3, padding=1)self.convolution4 = torch.nn.Conv2d(256, 256, 3, padding=1)self.pooling3 = torch.nn.MaxPool2d((1, 2), stride=(2, 1)) # notice stride of the non-square poolingself.convolution5 = torch.nn.Conv2d(256, 512, 3, padding=1)self.BatchNorm1 = torch.nn.BatchNorm2d(512)self.convolution6 = torch.nn.Conv2d(512, 512, 3, padding=1)self.BatchNorm2 = torch.nn.BatchNorm2d(512)self.pooling4 = torch.nn.MaxPool2d((1, 2), stride=(2, 1))self.convolution7 = torch.nn.Conv2d(512, 512, 2)def forward(self, x):x = F.relu(self.convolution1(x), inplace=True)x = self.pooling1(x)x = F.relu(self.convolution2(x), inplace=True)x = self.pooling2(x)x = F.relu(self.convolution3(x), inplace=True)x = F.relu(self.convolution4(x), inplace=True)x = self.pooling3(x)x = self.convolution5(x)x = F.relu(self.BatchNorm1(x), inplace=True)x = self.convolution6(x)x = F.relu(self.BatchNorm2(x), inplace=True)x = self.pooling4(x)x = F.relu(self.convolution7(x), inplace=True)return x # b*512x1x16class RNN(torch.nn.Module):def __init__(self, class_num, hidden_unit):super(RNN, self).__init__()self.Bidirectional_LSTM1 = torch.nn.LSTM(512, hidden_unit, bidirectional=True)self.embedding1 = torch.nn.Linear(hidden_unit * 2, 512)self.Bidirectional_LSTM2 = torch.nn.LSTM(512, hidden_unit, bidirectional=True)self.embedding2 = torch.nn.Linear(hidden_unit * 2, class_num)def forward(self, x):x = self.Bidirectional_LSTM1(x) # LSTM output: output, (h_n, c_n)T, b, h = x[0].size() # x[0]: (seq_len, batch, num_directions * hidden_size)x = self.embedding1(x[0].view(T * b, h)) # pytorch view() reshape as [T * b, nOut]x = x.view(T, b, -1) # [16, b, 512]x = self.Bidirectional_LSTM2(x)T, b, h = x[0].size()x = self.embedding2(x[0].view(T * b, h))x = x.view(T, b, -1)return x # [16,b,class_num]# output: [s,b,class_num]class CRNN(torch.nn.Module):def __init__(self, class_num, hidden_unit=256):super(CRNN, self).__init__()n = torch.nn.Sequential()n.add_module('vgg_16', Vgg_16())self.rnn = torch.nn.Sequential()self.rnn.add_module('rnn', RNN(class_num, hidden_unit))def forward(self, x):x = n(x)b, c, h, w = x.size()# print(x.size()): b,c,h,wassert h == 1 # "the height of conv must be 1"x = x.squeeze(2) # remove h dimension, b *512 * widthx = x.permute(2, 0, 1) # [w, b, c] = [seq_len, batch, input_size]# x = x.transpose(0, 2)# x = x.transpose(1, 2)x = self.rnn(x)return x
损失函数设计
刚刚完成了CNN层和RNN层的设计，现在开始设计转录层，即将RNN层输出的结果翻译成最终的识别文字结果，从而实现不定长的文字识别。pytorch没有内置的CTC loss，所以只能去Github下载别人实现的CTC loss来完成损失函数部分的设计。安装CTC-loss的方式如下：
git clone /SeanNaren/warp-ctc.gitcd warp-ctcmkdir build; cd buildcmake ..makecd ../pytorch_binding/python setup.py installcd ../buildcp libwarpctc.so ../../usr/lib
待安装完毕后，我们可以直接调用CTC loss了，以一个小例子来说明ctc loss的用法。
import torchfrom warpctc_pytorch import CTCLossctc_loss = CTCLoss()# expected shape of seqLength x batchSize x alphabet_sizeprobs = torch.FloatTensor([[[0.1, 0.6, 0.1, 0.1, 0.1], [0.1, 0.1, 0.6, 0.1, 0.1]]]).transpose(0, 1).contiguous()labels = torch.IntTensor([1, 2])label_sizes = torch.IntTensor([2])probs_sizes = torch.IntTensor([2])probs.requires_grad_(True) # tells autograd to compute gradients for probscost = ctc_loss(probs, labels, probs_sizes, label_sizes)cost.backward()
CTCLoss(size_average=False, length_average=False)# size_average (bool): normalize the loss by the batch size (default: False)# length_average (bool): normalize the loss by the total number of frames in the batch. If True, supersedes size_average (default: False)forward(acts, labels, act_lens, label_lens)# acts: Tensor of (seqLength x batch x outputDim) containing output activations from network (before softmax)# labels: 1 dimensional Tensor containing all the targets of the batch in one large sequence# act_lens: Tensor of size (batch) containing size of each output sequence from the network# label_lens: Tensor of (batch) containing label length of each example
从上面的代码可以看出，CTCLoss的输入为[probs, labels, probs_sizes, label_sizes]，即预测结果、标签、预测结果的数目和标签数目。那么我们仿照这个例子开始设计CRNN的CTC LOSS。
preds = net(image)preds_size = Variable(torch.IntTensor([preds.size(0)] * batch_size)) # preds.size(0)=w=16cost = criterion(preds, text, preds_size, length) / batch_size # 这里的length就是包含每个文本标签的长度的list，除以batch_size来求平均losscost.backward()
网络训练设计
接下来我们需要完善具体的训练流程，我们还写了个trainBatch函数用于bacth形式的梯度更新。
def trainBatch(net, criterion, optimizer, train_iter):data = train_iter.next()cpu_images, cpu_texts = databatch_size = cpu_images.size(0)lib.dataset.loadData(image, cpu_images)t, l = converter.encode(cpu_texts)lib.dataset.loadData(text, t)lib.dataset.loadData(length, l)preds = net(image)#print("preds.size=%s" % preds.size)preds_size = Variable(torch.IntTensor([preds.size(0)] * batch_size)) # preds.size(0)=w=22cost = criterion(preds, text, preds_size, length) / batch_size # length= a list that contains the len of text label in a batchnet.zero_grad()cost.backward()optimizer.step()return cost
整个网络训练的流程如下：CTC-LOSS对象->CRNN网络对象->image,text,len的tensor初始化->优化器初始化，然后开始循环每个epoch，指定迭代次数就进行模型验证和模型保存。CRNN论文提到所采用的优化器是Adadelta，但是经过我实验看来，Adadelta的收敛速度非常慢，所以改用了RMSprop优化器，模型收敛速度大幅度提升。
criterion = CTCLoss()net = Net.CRNN(n_class)print(net)net.apply(lib.utility.weights_init)image = torch.FloatTensor(Config.batch_size, 3, Config.img_height, Config.img_width)text = torch.IntTensor(Config.batch_size * 5)length = torch.IntTensor(Config.batch_size)if cuda:net.cuda()image = image.cuda()criterion = criterion.cuda()image = Variable(image)text = Variable(text)length = Variable(length)loss_avg = lib.utility.averager()optimizer = optim.RMSprop(net.parameters(), lr=Config.lr)#optimizer = optim.Adadelta(net.parameters(), lr=Config.lr)#optimizer = optim.Adam(net.parameters(), lr=Config.lr,#betas=(Config.beta1, 0.999))for epoch in range(Config.epoch):train_iter = iter(train_loader)i = 0while i < len(train_loader):for p in net.parameters():p.requires_grad = Truenet.train()cost = trainBatch(net, criterion, optimizer, train_iter)loss_avg.add(cost)i += 1if i % Config.display_interval == 0:print('[%d/%d][%d/%d] Loss: %f' %(epoch, Config.epoch, i, len(train_loader), loss_avg.val()))loss_avg.reset()if i % Config.test_interval == 0:val(net, test_dataset, criterion)# do checkpointingif i % Config.save_interval == 0:torch.save(net.state_dict(), '{0}/netCRNN_{1}_{2}.pth'.format(Config.model_dir, epoch, i))
训练过程与测试设计
下面这幅图表示的就是CRNN训练过程，文字类别数为6732，一共训练20个epoch，batch_Szie设置为64，所以一共是51244次迭代/epoch。
在迭代4个epoch时，loss降到0.1左右，acc上升到0.98。
接下来我们设计推断预测部分的代码，首先需初始化CRNN网络，载入训练好的模型，读入待预测的图像并resize为高为32的灰度图像，接着讲该图像送入网络，最后再将网络输出解码成文字即可输出。
import timeimport torchimport osfrom torch.autograd import Variableimport lib.convertimport lib.datasetfrom PIL import Imageimport as Netimport alphabetsimport sysimport Configos.environ['CUDA_VISIBLE_DEVICES'] = "4"crnn_model_path = './bs64_model/netCRNN_9_48000.pth'IMG_ROOT = './test_images'running_mode = 'gpu'alphabet = alphabets.alphabetnclass = len(alphabet) + 1def crnn_recognition(cropped_image, model):converter = lib.convert.strLabelConverter(alphabet) # 标签转换image = cropped_image.convert('L') # 图像灰度化### Testing images are scaled to have height 32. Widths are# proportionally scaled with heights, but at least 100 pixelsw = int(image.size[0] / (280 * 1.0 / Config.infer_img_w))#scale = image.size[1] * 1.0 / Config.img_height#w = int(image.size[0] / scale)transformer = lib.dataset.resizeNormalize((w, Config.img_height))image = transformer(image)if torch.cuda.is_available():image = image.cuda()image = image.view(1, *image.size())image = Variable(image)model.eval()preds = model(image)_, preds = preds.max(2)preds = preds.transpose(1, 0).contiguous().view(-1)preds_size = Variable(torch.IntTensor([preds.size(0)]))sim_pred = converter.decode(preds.data, preds_size.data, raw=False) # 预测输出解码成文字print('results: {0}'.format(sim_pred))if __name__ == '__main__':# crnn networkmodel = Net.CRNN(nclass)# 载入训练好的模型，CPU和GPU的载入方式不一样，需分开处理if running_mode == 'gpu' and torch.cuda.is_available():model = model.cuda()model.load_state_dict(torch.load(crnn_model_path))else:model.load_state_dict(torch.load(crnn_model_path, map_location='cpu'))print('loading pretrained model from {0}'.format(crnn_model_path))files = sorted(os.listdir(IMG_ROOT)) # 按文件名排序for file in files:started = time.time()full_path = os.path.join(IMG_ROOT, file)print("=============================================")print("ocr image is %s" % full_path)image = Image.open(full_path)crnn_recognition(image, model)finished = time.time()print('elapsed time: {0}'.format(finished - started))
识别效果和总结
首先我从测试集中抽取几张图像送入模型识别，识别全部正确。
我也随机在一些文档图片、扫描图像上截取了一段文字图像送入我们该模型进行识别，识别效果也挺好的，基本识别正确，表明模型泛化能力很强。
我还截取了增值税扫描发票上的文本图像来看看我们的模型能否还可以表现出稳定的识别效果：
这里做个小小的总结：对于端到端不定长的文字识别，CRNN是最为经典的识别算法，而且实战看来效果非常不错。上面识别结果可以看出，虽然我们用于训练的数据集是自己生成的，但是我们该模型对于pdf文档、扫描图像等都有很不错的识别结果，如果需要继续提升对特定领域的文本图像的识别，直接大量加入该类图像用于训练即可。CRNN的完整代码可以参考我的Github。
