Image Saliency Detection Based On Multi-graph Prior And Multilevel Features Connection Network

Image saliency detection refers to extracting regions of human interests from the image by using a computer to simulate the visual attention mechanism of the human eye.Image saliency detection methods are divided into models of traditional manual features and models of depth features.The traditional models are based on the prior knowledge of the existing dataset,and the saliency map produced cannot achieve satisfactory results in complex scenarios.With the rapid development of artificial intelligence,models of traditional manual features are gradually transformed into models based on deep features.Convolutional Neural Networks used in models based on deep features mainly uses neural networks to extract deep features that contain rich semantic information.Although saliency methods have made significant progress,the use of convolutional networks will lose the position and details of the object.In low-contrast images,misleading information may be introduced when detecting salient objects,resulting in incomplete object detection and background noise and unclear boundary prediction.In order to solve this problem,this paper proposes a saliency detection algorithm based on multi-graph model and multi-level features connection network.The algorithm performs saliency detection by combining traditional models and depth models.First,superpixel segmentation is performed on the image,and a K-Nearest Neighbor Graph and a K regular graph are constructed on the segmented image.On the KNN graph model,the salient values are calculated using the manifold sorting algorithm that uses the optimized boundary superpixels as query nodes.One the K regular graph model,the salient values are calculated according to the influence factors between superpixels.Then,the saliency values on the two models are fused to get the initial saliency map.The initial saliency map is used as prior information,and the original map is used as the input of the multi-level features connection network.The context information extraction module(CIEM)and attention model are used to extract the features of each layer.Finally,the multi-level features are used to connect the network to get the final saliency map.On three public datasets including ECSSD,HKU-IS,and PASCAL-S,the proposed algorithm is compared with the existing 14 algorithms.The experimental results show that the proposed algorithm detects complex scenes,especially in low contrast scenes.The salient object is more complete,and the saliency map is consistently highlighted inside.In order to further verify the performance of the algorithm in this paper,the saliency maps of this algorithm and several image saliency algorithms are applied to the field of image segmentation,and comparative experiments are performed on the ECSSD,HKU-IS,and PASCAL-S datasets.The experimental results show that the proposed algorithm which is used in segmentation of the saliency map can obtain salient objects more accurately.