A Binary Convolutional Encoder-decoder Network for Real-time Natural Scene Text Processing
This addresses the problem of memory and run-time inefficiency for real-time text processing on resource-constrained devices like advanced driver assistance systems, representing a strong specific gain.
The paper tackles real-time natural scene text processing by developing a binary convolutional encoder-decoder network (B-CEDNet) that achieves 90% and 91% pixel-wise accuracy on ICDAR datasets, with inference run-time of 4.59 ms and a network size of 2.14 MB.
In this paper, we develop a binary convolutional encoder-decoder network (B-CEDNet) for natural scene text processing (NSTP). It converts a text image to a class-distinguished salience map that reveals the categorical, spatial and morphological information of characters. The existing solutions are either memory consuming or run-time consuming that cannot be applied to real-time applications on resource-constrained devices such as advanced driver assistance systems. The developed network can process multiple regions containing characters by one-off forward operation, and is trained to have binary weights and binary feature maps, which lead to both remarkable inference run-time speedup and memory usage reduction. By training with over 200, 000 synthesis scene text images (size of $32\times128$), it can achieve $90\%$ and $91\%$ pixel-wise accuracy on ICDAR-03 and ICDAR-13 datasets. It only consumes $4.59\ ms$ inference run-time realized on GPU with a small network size of 2.14 MB, which is up to $8\times$ faster and $96\%$ smaller than it full-precision version.