論文筆記：An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application

阿新 • • 發佈：2018-12-31

1.歷史方法 1）基於字元的DCNN,比如photoOCR.單個字元的檢測與識別。要求單個字元的檢測器效能很強，crop的足夠好。 2）直接對圖片進行分類。9萬個單詞，組合成無數的單詞，無法直接應用 3）RNN,訓練和測試均不需要每個字元的位置。但是需要預處理，從圖片得到特徵序列，獨立做的，無法端到端。 2.propose method 1)端到端 2）任意長度 3）模型尺寸小 4）不針對特定字典，有無字典表現都很好 3.network如圖圖：

1）輸入圖片高度要一致 2）（創新性的地方）map2seq：在最後一個卷基層輸出的所有通道上，從左到右逐列拼接，得到特徵序列。按照寬度為1來輸入到LSTM，這樣可能多個長條特徵才會cover一個單詞。圖：

3）bp時需要seq2map，把梯度再拼成map 4）BLSTM利用雙向資訊，並stacked更深 5）標註序列的概率，通過CTC Layer,label不用具體到每個字元。公式圖：

任意一個label序列的概率 = 它的不同對齊方式的概率之和。主要是空格以及重複字母的影響 note:注意對映的時候是先刪除重複的字母，後刪除空格。示意圖：“AAA” … “AA”是LSTM輸出的不同預測的一個分佈。

6)優化方法採用adadelta。SGD似乎不work。 7）trick:長方形池化,BN end

論文筆記：An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application

1.歷史方法 1）基於字元的DCNN,比如photoOCR.單個字元的檢測與識別。要求單個字元的檢測器效能很強，crop的足夠好。 2）直接對圖片進行分類。9萬個單詞，組合成無數的單詞，無法直接應用 3）RNN,訓練和測試均不需要每個字元的位置。但是需要預處理，從圖片得到特

深度學習論文翻譯解析（二）：An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition

論文標題：An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition 論文作者： Baoguang Shi, Xiang B

論文筆記：An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application

論文筆記：An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application

深度學習論文翻譯解析（二）：An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition

《An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its...》論文閱讀之CRNN

論文筆記：雙線性模型《Bilinear CNN Models for Fine-Grained Visual Recognition》

論文筆記：Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling

論文筆記：DRAW: A Recurrent Neural Network For Image Generation

簡潔版 CVPR-2017論文筆記《Active Convolution: Learning the Shape of Convolution for Image Classification》

論文筆記《Active Convolution: Learning the Shape of Convolution for Image Classification》

論文閱讀筆記之——《DN-ResNet: Efficient Deep Residual Network for Image Denoising》

MSCNN論文解讀-A Unified Multi-scale Deep Convolutional Neural Network for Fast Object Detection

論文筆記|Towards End-to-End Lane Detection: an Instance Segmentation

【論文筆記】An End-to-End Model for QA over KBs with Cross-Attention Combining Global Knowledge

2017-ICLR End-To-End Optimized Image Compression論文筆記

深度學習筆記1：end-to-end、anchor box解釋、人體檢測程式碼

論文筆記（1）DenseBox: Unifying Landmark Localization with End to End Object Detection

【論文筆記07】End-To-End Memory Networks

論文翻譯：Generalized end-to-end loss for speaker verification

Overview：end-to-end深度學習網絡在超分辨領域的應用（待續）

【USE】《An End-to-End System for Automatic Urinary Particle Recognition with CNN》

《End-to-End Learning of Motion Representation for Video Understanding》論文閱讀

論文筆記：An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application

相關推薦