Analysis and Recognition of Hand-Drawn Images with Effective Data Handling

Everyday data is being produced at a mind-boggling rate and at current pace, we are generating more than 2.5 quintillion bytes of data each day. With more data, there comes the chance of generating and understanding more useful information out of it. With this view in mind, we are presenting our work on Quick Draw dataset. It is a repository of approximately more than 50 million hand-drawn drawings of about 345 different objects, contributed by over 15 million users. The work we have presented is a useful approach to perform a country-wise analysis of the styles that people use while making strokes to draw an object or image. Since we are dealing with a huge dataset, we are also presenting a powerful technique to effectively reduce the data and its dimension which results in much more simplified and concise data. Based on the simplified data, we are enhancing the performance of our models to recognize these hand-drawn objects based on the strokes that a user has used for drawing them. For the recognition of objects, we have worked on traditional machine learning models - Nearest Neighbor (K-NN), Random Forest Classifier (RFC), Support Vector Classifier (SVC) and Multi-Layer Perceptron model (MLP) by selecting the best hyper-parameters. To make the recognition results better, we have also presented our work on deep learning models - CNN, LSTM, and WaveNet networks.