关键词: 图像分割, 注意力, 特征图, 场景-目标显著图, 跨尺度融合

Abstract: An attention-based approach for image segmentation is proposed. It integrates the bottom-up and top-down attention mechanism, to form a scene-target selection method for the target objects in an image. In the multi-scale space of image, this algorithm simultaneously extracts the intensity, color and orientation features of the scene image and the color feature of the target object to generate the scene-target saliency map. Then, it processes the saliency map by combination the multi-scale scene-target images with normalization of the image features. Finally, the target object is obtained by double-interpolation and black-white segmentation of the scene image. By applying the algorithm to the images in natural scene and indoor environment, experiment is conducted. The experimental results indicate that the algorithm can successfully segment the scene image and extract the target object in any condition, and exhibit good robustness even for the scene image with noisy objects.

Key words: image segmentation, attention-based, feature map, scene-target saliency map, multi-scale combinations