Research On Image Saliency Detection Model Based On Deep Learning

Deep learning is a learning model based on large-scale data sets and deep structures.With the development of the times,the computing power of computers has been greatly improved,and the number of large-scale data sets has increased day by day,providing a realistic basis for deep learning.Deep learning extracts the deep abstract features of data through the stacking of network layers,which is impossible for traditional machine learning methods.Today,deep learning has made outstanding achievements in computer vision,natural language processing and other fields.Human’s innate visual system can quickly detect prominent areas in the visual scene.This ability to detect salient objects plays an important role in the human visual system.It enables humans to focus their limited perception and analysis capabilities on the most important areas of the scene,so as to quickly extract key information.The analysis and simulation of this saliency object detection ability is a basic research direction in the field of computer vision.Saliency detection helps to quickly find the most representative objects or regions in the scene,and provides support for some complex visual problems.With the development of deep learning,saliency detection methods based on deep learning are also emerging.This paper studies the image saliency detection model based on depth learning.However,there are still many deficiencies in the current saliency detection methods.In terms of network structure,the existing mainstream saliency detection methods still have some defects.Different from the early saliency detection methods based on a single mode,most existing methods choose to extract features from images of multiple modes,which can provide multi-dimensional information for the detection of salient targets.But when fusing the features of different modes,the existing methods usually adopt symmetrical structure,treat the features of the two modes equally,and do not consider the interference of some unreliable thermal images on model reasoning.When aggregating different levels of features,the current mainstream encoder-decoder structure uses a continuous cascade structure,resulting in the dilution of high-level semantic information in the process of continuous aggregation,affecting the accuracy of model reasoning.In the aspect of feature enhancement,the existing mainstream saliency detection methods are not deep enough.For-high-level features,existing methods usually rely on the convolution layer stacked in the backbone network to extract the context information,but this method has a slow growth of feature receptive field,and the context information is not comprehensive enough.For low-level features,existing methods lack further mining of texture information.In view of the shortcomings of the existing methods in the above two aspects,this paper studies two image saliency detection models based on depth learning respectively.The main work and innovation are as follows:In view of the shortcomings of existing methods in network structure,this paper proposes a saliency detection model based on cross-modal feature fusion and semantic information flow supplement.The experimental results show that the detection effect of this model on three datasets is better than other existing methods in the field.In this model:1.Aiming at the problem of thermal image reliability ignored by the cross-modal feature fusion structure used in the existing methods,this paper proposes a cross-modal feature fusion method based on thermal image reliability.This method provides weight for the cross-modal fusion of thermal features and RGB features by reasoning the reliability of thermal images,thus avoiding the introduction of noise from some unreliable infrared images,and fully fusing reliable thermal features and RGB features.2.Aiming at the problem of semantic information dilution when existing methods aggregate different levels of features,this paper proposes a decoding method based on semantic information flow supplement.The existing saliency detection models mainly use the encoder-decoder structure with skip connection.Multiple decoding of this structure will lead to dilution of high-level semantic information.This method makes up for the defect of this structure by extracting semantic information from highlevel features and adding explicit information flow.In view of the shortcomings of existing methods in feature enhancement,this paper proposes a saliency detection model based on context information enhancement and texture feature mining.Experimental results show that the detection effect of this model on three data sets is better than other existing methods in the field.In this model:1.Aiming at the problem that the existing methods are not sufficient to extract context information when obtaining high-level features,this paper proposes a context information reinforcement method based on graph convolution network.The existing saliency detection methods mainly obtain the context information through the stack of the convolution layer,and the effect is limited.This method makes use of the feature that graph convolution network can exchange information globally,and more effectively extracts context information from the original features.2.Aiming at the lack of further mining of texture features when existing methods acquire low-level features,this paper proposes a texture feature enhancement based on statistics and quantification.The existing saliency detection methods focus on the extraction of high-level features,but pay little attention to low-level features.In this method,the texture features in the low-level features are enhanced by statistical quantification.