Research On Visual Feature Space Perception Algorithms Based On Deep Learning

In recent years,with the development of artificial intelligence technology,computer vision has made great progress as its main application field.Faced with the explosive growth of image data,for different vision tasks how to visualize spatial feature perception to complete the corresponding task goals,becoming a research hotspot in the field of computer vision,and is widely used in many computer vision tasks,such as object detection,semantic segmentation,style transfer and so on.When the computer performs the visual processing task,the perception ability of the object depends on the ability of the extracted feature to represent the object,and the extracted task needed complete visual feature space information,which plays an important role for the computer vision task.Therefore,the study of spatial perception of visual features has the important theoretical value.Deep learning represents the image layer by layer by extracting higher order information hierarchically.This representation way effectively simulates the human visual perception system and the neuron transfer process,which can construct feature representations that are more efficient than traditional feature perception methods.Therefore,in order to solve the problem of in learning and mapping transfer of feature space in visual feature perception,in this paper,combined with the advantages of deep learning,the following research are proposed focusing on the three problems of material visual feature perception,saliency feature perception and feature adaptive perception(?):(1)Aiming at the problem of the mutual mapping of object feature space in material visual feature perception,this paper proposes an visual feature mapping algorithm based on Cycle-GAN for natural scene image.Firstly,we obtain the reflection layer from the image represented the important features of the material vision.Then,we get the image by segmenting foreground and background image from the reflection layer image.Finally,we obtain a high-order expression of the visual feature space of the image object material using Cycle-GAN to perform unsupervised learning on the material visual feature,and realize the visual feature mapping of the object material for the natural scene image.The experimental results show that the proposed algorithm can effectively obtain the material visual feature of the natural scene image object and perform the material visual feature mapping,which has been achieved better subjective and objective experimental effects.(2)Aiming at the problem of the difference representation measure between the object and background feature space,this paper proposes a salient object detection algorithm based on high-low-level feature fusion.Firstly,we extract the low-order feature space information making use of the role of the traditional methods in salient object detection.Then,we extract the high-order saliency feature using the advantages of the deep learning method in the feature extraction process,and obtain the saliency feature space of the object using the side-output method to complete the fusion of the feature.Finally,in the post-processing process,the boundary information is optimized to obtain a more complete saliency area to achieve salient object detection.The experimental results show that the proposed algorithm can make full use of the advantages of high-low-level order feature,and effectively combined the high-lowlevel order feature information,and comparative experiments show that the accuracy of salient object detection is better.(3)Aiming at the problem of adaptive mapping of bbject and background feature space,this paper proposes an object adaptive hiding algorithm based on feature space statistical information mapping.Firstly,based on the salient object detection method in(2),the object area and the background area are divided.Then,using convolution calculation,the background area and the object area are coded respectively and mined the hidden high-order representation to the spatial statistical information representing the background style features mapped into the object feature space,and achieve the effect of feature hiding using the feature space fusion to weaken the original features.Finally,object adaptive hiding is achieved by image fusion after feature mapping in the object and original background.The experimental results show that the proposed algorithm can achieve object adaptive hide according to different background environments,and has better subjective visual effects.At the same time,through the inverse verification of the salient object detection,it can be shown from the objective data that the proposed algorithm can effectively reduce the saliency for the salient objects in images,which has a better hidden effect.