Research On Cross-Level Saliency Detection Based On High-Level Semantic Feature

Saliency detection refers to the process of finding the most "attractive" object or objects from a picture,which is essentially a computer reproduction of psychological process.Deep convolution neural network technology has shown incredible ability in the application of various fields of graphics and images,and it is no exception in the field of saliency detection.At present,the most advanced methods in the field of saliency detection are based on deep convolution neural network technology.Lee put forward a uniform coding scheme called ELD(Encoded Low Level Distance)algorithm based on low-level and high level features,which is worth studying.Referring to the core idea of ELD algorithm,this paper fully integrates the low-level detail features with high-level semantic features of VGG16 network in multi-level and cross-level,and adds the spatial and channel attention modules in order to enhance the performance.In this paper,a cross-level saliency detection deep neural network is designed and implemented.After testing on up to 11 popular public data sets,it shows that the algorithm put forward in this paper is effective and have good results.The innovation of this paper is as follows:1.The architecture of ELD model is improved.In this paper,the low-level features in ELD are replaced by the low-level feature map in VGG16,and are integrated with the high-level feature map for learning.2.A new network interconnection structure is proposed.In the ELD model,neural network is only used as a tool to extract high-level features,without considering the connectivity between the layers.In this paper,each layer of neural network is connected and fused to achieve the purpose of feature fusion at different levels.3.The weight sensitivity of space and channel is enhanced.Spatial and channel attention mechanism(SCSE)is introduced into the process of hierarchical interconnection to optimize the weight of each level.Although the training speed is a little slow,it achieves better results.