| Saliency detection has led to many studies in the field of computer vision and has produced many applications.Saliency detection aims to extract salient areas(i.e.,areas of human interest)from pictures or videos.Traditional saliency detection methods fail to distinguish salient objects in images presenting complex scenes.With the successful application of neural networks in computer vision,more and more deep learning based saliency detection methods are applied to detect salient objects,and deep learning based saliency detection methods generate better saliency detection results than traditional saliency detection methods.This paper proposes fusing the results of the traditional heuristic saliency detection methods to assist the deep model to train for better prediction results.By proposing to use the salient detection map of the traditional methods as priors in a neural network,this paper aims to obtain a more complete and accurate saliency detection map both in simple scenes and complex scenes.Although there are many traditional heuristic saliency methods,it is not effective to combine the results of all methods.Therefore,by considering the speed and F-measure values of traditional saliency detection methods,we propose a strategy for selecting suitable traditional saliency detection methods.The training is often accompanied by a large amount of training data.For the saliency detection problem,this paper uses a data augmentation method which differs from other classification problems,and crops image while maintaining the integrity of the salient object.Finally,we improve the VGG-16 network structure for feature extraction and saliency map genaration.Comparing with some existing traditional saliency detection methods and some deep learning based saliency detection methods in ASD datasets,ECSSD datasets,PASCAL-S datasets and SED2 datasets,both qualitative and quantitative analysis of results show our algorithm proposed in this paper is able to make better significant predictions. |
PDF Full Text Request