CCC: Weakly Supervised Semantic Segmentation using Class Contribution Classification

We propose a new semantic Segmentation technique that depends on the convolutional neural network of Xception using a modified categorical cross-entropy function to perform class contribution classification. We introduce a class contribution function for each class contribution in the image instead of categorical cross-entropy for maximum class classification. This approach is to be used for weakly supervised semantic segmentation (WSSS) using image-level annotation instead of the pixel-level annotations used in fully supervised semantic segmentation. We trained and tested our approach on both PASCAL VOC2007 and VOC2012 datasets. We show that our approach outperformed many other weakly supervised methods, specifically, it can attain a segmentation accuracy of 49.7% on PASCAL VOC2007 test set and an accuracy of 41.3% on PASCAL VOC2012 test set while the network is trained to learn the semantics of the objects in the image using the cheap image label annotation instead of the expensive and time-consuming segmentation masks.