Salient Object Detection: Integrate Salient Features in the Deep Learning Framework

Salient object detection in complex environments brings the challenge from the collections of large number of training images for deep learning algorithm. It is difficult to collect the enough number of training data for varied salient objects in different scenes, and furthermore the salient objects are usually compared with the background. This paper proposes a novel method to integrate the salient features into the deep learning framework, and design a parallel multi-scale structure of the neural network to enhance the ability to detect salient objects. In addition, a multiple stage method is proposed to optimize the training process and make the datasets more prominent in the commonality of the salient features, which effectively reduces the difficulty of correctly distinguishing salient objects in complex environments. Experiments show that the proposed approach enhances the ability of neural networks to learn specified features and improves the detection effect of salient objects in complex scenes.