Research On Salient Object Detection Based On Generative Adversarial Networks

In recent years,with the popularization of smart terminal devices,image data have exploded,and the efficient and effective acquisition of useful information in images will greatly improve our work efficiency.Because of the visual attention mechanism,human beings will first focus on the objects we are interested in when facing a scene or image,and ignore the background information,which will help improve the information processing rate of our brain.Salient object detection is the use of intelligent algorithms to simulate this attention mechanism,allowing the computer to automatically identify and segment the salient objects in the image.Salient object detection has a wide range of applications in the fields of automatic driving,surveillance security,and image retrieval,so its research value is very high.In this paper,some problems in the field of salient object detection are studied,and two saliency detection models are proposed: the edge constraint based salient object detection model and the salient object detection model based on generative adversarial networks.The network structure of the edge constraint based salient object detection model is an improved U-net.It can fully fuse the deep semantic information of the image with the shallow features.We propose a novel loss function with an edge constraint term which is based on the idea of image convolution to extract the target edge.The model solves the problem of blurring of the object edge in the current salient object detection and the inconsistent prediction results of each region within the object.Our experiment was conducted on five benchmark data sets,three of which were test sets,and 11 models were selected for comparison.Experimental results show that the saliency map predicted by our model is superior to other models in object edge and detection accuracy.On the DUT-OMRON dataset,the F-measure of our model is 3.77% higher than that of the comparison model.The salient object detection models based on deep learning rely on a specific loss function during training,and image information that cannot be reflected in these functions is often not utilized.Many salient object detection models(especially traditional models)are only applicable under certain conditions,often with limitations on the size and number of salient objects in the input image.In order to solve the above problems,we propose a salient object detection model based on generative adversarial networks.We use conditional generative adversarial networks to learn a mapping from the natural image domain to the saliency map domain,in which the generator outputs the most realistic saliency map as it is trained,and the discriminator separates it from the ground-truth as far as possible.The discriminator will comprehensively consider various information of the image when discriminating the true and false saliency maps,and the discriminant score will guide the training of the generator,so the discriminator is equivalent to a part of the generator loss function.In addition,according to the idea that the matrix and its own multiplication can expand the difference information between the matrices,we also propose a new spatial constraint term.Our experiments are performed on six benchmark datasets and 11 state-of-the-art models are selected.The experimental results show that the generative adversarial networks can also achieve excellent performance in the salient object detection task.