This paper proposes a discriminative regional feature integration approach for salient object detection. The method formulates saliency map computation as a regression problem, using multi-level image segmentation and supervised learning to map regional feature vectors to saliency scores. Saliency scores across multiple layers are fused to produce the final saliency map. The key contributions are: (1) a discriminative regional feature integration approach that automatically integrates high-dimensional regional saliency features and selects discriminative ones; and (2) superior performance on six benchmark datasets compared to state-of-the-art methods. The method is efficient, running as fast as existing algorithms. The approach uses three types of regional saliency features: regional contrast, regional backgroundness, and regional property. These features are combined using a Random Forest regressor to predict saliency scores. The method is evaluated on five benchmark datasets (MSRA-B, iCoSeg, SED, ECSSD, DUT-OMRON) and outperforms existing methods in terms of AUC, PR, and ROC curves. The approach is also robust to challenging cases, such as when salient objects are near the image border or in cluttered scenes. The method is efficient, with training time around 24 hours and inference time comparable to existing methods. The paper also discusses the advantages of supervised learning over heuristic methods, including better generalization and automatic feature integration. The approach is limited in cluttered scenes where regional contrast and backgroundness features are not sufficient.This paper proposes a discriminative regional feature integration approach for salient object detection. The method formulates saliency map computation as a regression problem, using multi-level image segmentation and supervised learning to map regional feature vectors to saliency scores. Saliency scores across multiple layers are fused to produce the final saliency map. The key contributions are: (1) a discriminative regional feature integration approach that automatically integrates high-dimensional regional saliency features and selects discriminative ones; and (2) superior performance on six benchmark datasets compared to state-of-the-art methods. The method is efficient, running as fast as existing algorithms. The approach uses three types of regional saliency features: regional contrast, regional backgroundness, and regional property. These features are combined using a Random Forest regressor to predict saliency scores. The method is evaluated on five benchmark datasets (MSRA-B, iCoSeg, SED, ECSSD, DUT-OMRON) and outperforms existing methods in terms of AUC, PR, and ROC curves. The approach is also robust to challenging cases, such as when salient objects are near the image border or in cluttered scenes. The method is efficient, with training time around 24 hours and inference time comparable to existing methods. The paper also discusses the advantages of supervised learning over heuristic methods, including better generalization and automatic feature integration. The approach is limited in cluttered scenes where regional contrast and backgroundness features are not sufficient.