Improved Image Semantic Segmentation Based on Cascade Data Augmentation

A large portion of the research for the task of image semantic segmentation focuses on optimizing deep learning networks to improve the state-of-the-art results. In this work, however, we take a different approach in order to enhance the performance of state-of-the-art methods for image semantic segmentation. Benefiting from the internal learning process of deep learning models we propose a Cascade Data Augmentation (CDA) technique that can effectively increases the size of data based on the characteristics of the dataset in hand. CDA utilizes basic image transformation techniques such as geometric and photometric transformations to augment the size of data. In particular, CDA blends these transformations by assigning two adjustable parameters that are 1) the probability of applying the transformation operation and 2) the magnitude of the operation. We evaluate the effectiveness of our method on the validation set of Pascal Context benchmark dataset for image semantic segmentation. Experimental results prove that incorporating CDA with baseline deep learning models can increase their performance. In particular we attain a average gain of ≈ 4.72 pixel accuracy and ≈ 1.9 mean intersection over union Pascal-Context dataset. Moreover, CDA is a generic method that can be integrated with any deep learning model and attains improved performance while adding trivial extra training time of the network.