Transfer Learning for Automatic Image Orientation Detection using Deep Learning and Logistic Regression

The number of images produced each day increased significantly. The ability to detect and correct an image's orientation can provide several advantages in computer vision. This paper presents a new framework based on a transfer learning technique for automatically detecting image orientation. To implement the power of deep neural networks, we applied a convolutional neural network model pre-trained on the ImageNet database for feature extraction. Then, we built a multi-class logistic regression classifier to detect the four image orientation probabilities corresponding to the following orientations (0 for no orientation, 90, 180, and 270). We tested our model on the SUN-397 dataset, one of the most extensive data sets currently used for image-orientation detection tasks. We conducted a cross-dataset evaluation for in-depth testing and analysis. We also examined our model using different old and recent state-of-the-art convolutional neural network (CNN) baselines. We demonstrate that our model yields promising results based on transfer learning for feature extraction combined with a one-vs-rest logistic regression classifier. Our proposed model surpassed the state-of-the-art results in terms of accuracy and performance.