Research On Image Saliency Region Detection Technology Based On Depth Learning

With the continuous improvement of people's living standards.The requirements for the accuracy and efficiency of computer image analysis are also increasing.People pay more attention to efficient and stable image recognition systems.Image saliency area detection technology has received more and more attention in the field of computer vision research.This paper introduces the related algorithms of deep learning networks based on existing algorithms.A more reasonable image saliency detection scheme is proposed,and then experiments and analysis are performed.In the research of this paper,the background,purpose and significance of the research and the research status of image saliency detection model are introduced.The related algorithms and theories are introduced,including the theoretical basis of visual saliency,image saliency characteristics,image saliency detection based on depth learning and image saliency evaluation criteria.Then innovatively proposed image saliency region detection based on multi-scale void convolution and contour learning.At present,in saliency detection,traditional algorithms focus more on the comprehensive application of multi-scale and multi-feature.The validity of features is verified by a large number of experiments.Many feature fusion algorithms are experimental selection of several features for fusion.After widely using the neural network model,feature selection is no longer a problem.By training a large number of sample sets and improving the network architecture,the neural network itself learns the features with good detection effect,which greatly improves the accuracy and robustness of the algorithm.However,the traditional neural network algorithm simply uses manual labeled binary graph(GT)to learn the foreground and background regions,and after training the network,the foreground regions can be detected directly.This often leads to some false detection in the background area.In this paper,a saliency region detection algorithm based on multi-scale void convolution and contour learning is proposed.The foreground region can be detectedmore accurately.Firstly,the neural network is trained by modifying the labeled binary graph and classifying the foreground,background and outline.By training the network to learn the edge features,the network can learn the foreground features more accurately.The hollow convolution is introduced as the convolution core,and the hollow convolution layer of multiple scales is used instead of the original full connection layer to better preserve the details of the original image.Through these two improvements,the network can better detect the foreground and finally get the salient map with high saliency value.Finally,the saliency region detection of full convolution network image based on segmentation is proposed.The saliency detection model based on convolution network is different from the traditional saliency detection method which extracts features manually.The convolution neural network enables the network to automatically extract effective features through the convolution layer,so that the saliency map can be calculated using effective features.Therefore,image saliency detection based on convolution neural network can get more effective features than traditional methods,so as to get better detection results.