A novel approach for indoor-outdoor scene classification using transfer learning

Abstract

Scene understanding and analysis has gained significant importance and widely used in computer vision and robotics field. Classification of complex scenes in a real-time environment is a difficult task to solve. Convolution Neural Networks (CNNs) is a widely used deep learning technique for the image classification. But the training of CNNs is not an easy task since it requires large scale datasets for training. Also, the construction of CNN architecture from the scratch is a complex work. The best solution for this problem is employing transfer learning which gives desired result with small scale datasets. A novel approach of Alexnet based transfer learning method for classifying images into their classes has been proposed in this paper. We selected 12 classes from publicly available SUN397 dataset out of which 6 are indoor classes and the remaining 6 are outdoor classes. The model is trained with indoor and outdoor classes separately and the results are compared. From the experimental results we found that the model exhibited the accuracy of 92% for indoor classes and 98% for outdoor classes.