A-bone1
A-bone1
@arcral 您好,谷歌给的官方源码是可以对多行文字进行识别的,但是我感觉这样识别结果可能不会比先检测裁取,再分别识别效果好
@aagainbeyond2 用的1.4版本
我训练时也是这样,刚开始全是空格,接着只能预测出相同的汉字,后面就好转了。请问你是多少数据集训练了几个step呢。同时,建议查看一下tensorboard,里面能呈现原始图像、真实文本和预测文本
@thachkysanh1996 This should be a problem with coding and decoding. You can check the relevant information. You can also use python2 to try the correctness of the result through this...
@thachkysanh1996 The type of task depends on your training data. If your training data is photo ocr, the model can work on photo ocr.
@thachkysanh1996 I have not tried this situation, but I think it is possible to add a text detection network in front of the text recognition network. It may be better...
@thachkysanh1996 Is the prediction result of the test set similar to the training set accurate? If the training set differs greatly from the test set style, the accuracy is really...
你好,我当时在训练模型时,前阶段的预测结果也只有相同的汉字。但随着训练的进行,模型就可以预测出正确的汉字了。所以,其中一个原因可能是训练次数不够,另一个原因可能是生成tfrecord时文本标签没有正确生成。建议可以打开tensorboard,看看image与text是否能够一一对应
训练中应该没用到,但是在tensorboard中会根据image/text呈现样本真实标签,建议可以用tensorboard打开logs中的文件进行详细内容的查看。其中,也能直接看到随机采样一些预测文本。
@wkhunter 你好,我当时180万的合成图片大概1个epoch就有初步效果了,后面loss也稳定在了40左右,但是,随着迭代次数的增多,验证集上的准确度是有在一直提升的,你可以在训练的同时运行在验证集上运行python eval.py,然后调节里面每次测试准确率的间隔时间,看准确率是否一直在持续增加