MSHSCNN: Multi-Scale Hybrid-Siamese Network to Differentiate Visually Similar Character Classes

摘要

We address the character recognition challenge of similar-looking character classes. Human Vision System often misinterprets visually similar characters while they are present at singular instances. Often, we pay soft attention to a combination of characters to read and interpret a word. A graphical readout of a display device dashboard shows characters generated from a custom vectored font library. A poorly defined character library or dimension adjustment of fonts in the pre-defined layout can often distort the geometric shape of characters. It requires additional intellectual cues for a human to understand and causes poor accuracy by an automated system responsible for character recognition. Even after generalizing over a large-scale dataset, a well-trained character recognition engine misclassifies these custom character images and exhibits low model confidence. In this paper, we optimize a multi-scale Siamese network using multitask Learning to learn significant discriminative features of visually similar characters from a few labeled samples of a custom vectored English font dataset. Using classification and similarity learning, Multitask Learning improves recognition performance and introduces strong inductive biases. Experiments show that our method can effectively distinguish visually similar characters and improves overall classification accuracy.