Analysis of Feature Extraction and Classification Models for Lip-Reading

Lip Reading is gaining momentum to be one of the toughest challenges in the Computer Vision society. We have analyzed feature extraction methods for accurately representing the lip contours in terms of feature representations and how these features were trained to give the output in the form of classes - phonemes, words, and sentences. Strong emphasis is given on the challenges that have motivated the creation of different feature extraction methods and neural network models specific to this task and how these challenges were solved by making modifications in the existing algorithms or using alternative algorithms. Feature extraction methods are made robust to variance in illumination, pose, etc. Once, we have these features we have dwelt more into the classification algorithms such as the Random Forest and the Support Vector Machine. Neural network models have been developed which aim to capture the spatial as well as the temporal features in the video datasets. We hope that the reader of this paper will be able to train better models after gaining an insight into how these challenges have been solved in the past.