Salient Region Detection by Integrating Intrinsic and Extrinsic Cues without Prior Information
Ji Ma, Jingjiao Li, Zhenni Li, Jiao Jiao · Journal of Engineering Science and Technology Review · 2017
In recent years, independence on prior information and the efficient integration of multiple visual cues have become an important topic in salient region detection.A salient region detection framework that integrates intrinsic and extrinsic visual cues was proposed in this study to remove the limitations of prior information.First, salient prior map was generated using statistical texture representations and by classifying superpixels into three classes of background, salient regions, and unknown regions.Second, the final saliency map was computed by integrating multi-channel color features and color contrast saliency factors based on self-generating salient prior map.Finally, the experiment was conducted on popular benchmark datasets, namely, Microsoft Research Asia (MSRA) and Extended Complex Scene Saliency Dataset (ECCSD), using two standard criteria: precision-recall rate and F-measure rate.Results demonstrate that the proposed method is more competitive than methods based on prior information, such as dense and sparse reconstruction (DSR), absorbing Markov chain (MC), and robust background detection (RBD).The proposed method achieves an 8% reduction in computing complexity compared with DSR.Compared with the learning-based method without prior information, the performance of the proposed method resembles that of discriminative regional feature integration (DRFI) and its complexity is reduced by 70%.This study provides a novel method to improve the performance of salient region detection and to avoid prior information.