Object Recognition in Hand Drawn Images Using Machine Ensembling Techniques and Smote Sampling

Abstract

With the increase in the advancement of technology, the size of multimedia content generated is increasing every day. Handling and management of this data to extract the patterns result in a more optimized and efficient way is the need of the hour. In this proposed work, we have presented techniques to classify the stroke-based hand-drawn object. We have used the Quick Draw dataset which is a repository of approximately 50 million hand-drawn drawings of 345 different objects. Our research presents an approach to the classification of these drawings created using hand strokes. We are converting the given raw image data, to much more simplified and concise data and then performed oversampling on data belonging to classes with fewer instances using Synthetic Minority Over-sampling Technique (SMOTE) method to balance the distribution of each class in the dataset. Finally, we are classifying the drawings using K-Nearest Neighbor (K-NN), Random Forest Classifier (RFC), Support Vector Classifier (SVC) and Multi-Layer Perceptron model (MLP) by working on their hyperparameters for the best-achieved classification result. The proposed solution attains the accuracy of 82.34% using best hyperparameter selection.