Charge Own Job: Saliency Map and Visual Word Encoder for Image-Level Semantic Segmentation