OCR with ctc_loss_layer to read captchas

带ctc损失层的keras OCR模型来读取网页上的captcha;

训练模型最后是ctc_layer, 训练后模型不能直接用于推断，另外建立推断模型，使用原基础模型的输入 和 最后Dense层的输出 而建立新模型。推断结果也需要进行额外的解析

in Mandarin/zh, the simplest way, keras Image Classification, build a model to classify cats and dogs, total two classes, that is, cat, dog. You download pics dataset of cats and dogs from microsoft, then apply augmentation ( translate, rotate, flip, contrast), build a residual convo model, compile and fit, save the model or keep the weights, now you can predict / inference if it is a cat or a dog. I took some pics of cats and dogs in the yard, nnnn, the result is not so good, I mean the accuracy is bad. Perhaps I shall let the cat/dog take major surface ratio like 80 -90% when shooting pics.

transfer learning in Mandarin: 迁移学习，用预训练好的权重， EfficientNet, 不要top, 只有权重，冻结权重， 再建立自己的top, 训练后， 再把 后20层 非BN层的 解冻，改为可以训练权重， 再训练10个epo, 达到令人满意的准确度，而耗用的资源/计算力 非常有限