A Deep Learning-Based Regression Scheme for Angle Estimation in Image Dataset

A machine needs to recognize orientation in an image to address various rotation related problems. To calculate this rotation, one must require the information about different objects that present into the image. Hence this becomes a pattern recognition task. By using Deep Learning this issue of calculation of image rotation can be addressed as deep learning possess excellent ability of feature extraction. This paper proposes a novel deep learning-based approach to estimate the angle of rotation very efficiently. Kaggle dataset (Rotated Coins) and Caltech256 has been used for this research, but the data available was limited hence this research utilize data augmentation by rotating the given dataset at random angles. Initially the unlabeled image has been rotated at different angles and store the values to be used as training dataset. Finally at the output a regression layer has been used to identify the angle of rotation for input image. The proposed deep learning approach provides a better result in terms of validation parameters like R-square, MSE, MAE. With proposed approach the value of R-square, MSE, and MAE for Kaggle dataset (Rotated Coins) obtained is 0.9846, 0.0013 and 0.0127 respectively. While forCaltech-256 Dataset proposed approach reportedR-square, MSE, and MAE of 0.9503, 0.0039 and 0.0240 respectively. The proposed approach also helps in finding the position of an object by calculating the angle of rotation in an image.