<p>In this paper, we propose a new approach for weakly supervised semantic segmentation, with only image-level labels. Recent methods follow a common pipeline consisting in generating pseudo-masks from class activation map (CAM) obtained from the feature given by the final layer of a classification network to train a segmentation network. To get accurate pseudo-masks, some approaches focus on enhancing CAM, while others aim to further improve the pseudo-masks using existing CAM. This paper focuses on the latter approach, which generally lacks the balance between semantics and resolution, and which is designed as a global process that does not focus on the singularity of each image but considers the whole training set. The proposed Multi-feature Image-specific Pixel classification approach combines semantics from deep features with high-resolution shallow features. It consists of three steps: the generation of seed from CAM, a sample selection process that provides a training set for a pixel-wise Support Vector Machine (SVM) classifier, and a self-enhanced SVM training and predicting. The potential of the method is demonstrated by evaluating its maximum performance limit, as a function of the selected shallow features and using the ground truth masks. Finally, experiments show that this approach outperforms the baseline method, and the performances are also comparable with recent state-of-the-art methods on both PASCAL VOC 2012 and MS COCO 2014 datasets.</p>