paper review: "Fast DenseNet: Towards Efficient and Accurate Text Recognition with Fast Dense Networks"

https://arxiv.org/ftp/arxiv/papers/1912/1912.07016.pdf

TLDR: key points

this paper proposes to use densent+CTC for text recognition, and in that process propose some modifications to original densenet which are

• 1. new block, called Fast Dense Block(FDB)
• 2. FDenseNet-U: fast densenet + upsampling block
• 3. use convolution layer with stride 2 instead of maxpooling
• 4. apply depth-wise separable convolution

Fast Dense Block

• combining good attributes of residual block, dense block, residual dense block.
• uses addition instead of concatenating when merging intermediate results in a single block. This helps to reduce computation. Concatenation layer is only used at the end of the block.

FDenseNet-U

• does upsampling to compensate for possibly lost information during downsampling.
• Then pass on the result to CTC for text recognition.

Comments

Unfortunately, since this paper was intented to target text recognition and not simple classification, there is no experiments done to compare classification performance using original densenet vs. Fast DenseNet.