Implementation of the method presented at "Efficient Learning-Free Keyword Spotting" (https://ieeexplore.ieee.org/abstract/document/8378004/)
A method for KWS, consisted of three steps:
- Preprocessing (contrast and main-zone normalization)
- Feature extraction (projections of Oriented Gradients - POG) over image zones
- Matching sequences of descriptors with dynamic programming
Implementation Highlights:
- Implemented using Matlab 2015a (possible compatibility issues for other versions).
- Feature extraction is using multiple threads (Matlab's parfor).
- The matching algorithm is implemented in C (inner_valid_seq_multi.c) and was compiled via mex for Linux arch64. For different architecture re-compile the mathing algorithm:
mex inner_valid_seq_multi.c
- The dataloaders for both the datasets of ICFHR 2014 Keyword Spotting Competition are provided (Bentham14 & Modern14).
- 3 methods are supported:
- 'Global' : simple holistic POG descriptor
- 'GlobalZoned' : use POG descriptor over zones across the word image
- 'Sequential': the proposed method using the sequential matching. The multi instance case, decribed at the manuscript, is also supported.