Research On Techniques For Saliency Detection Based On Deep Neural Network

Saliency detection is to extract the key target or region from images.As an important branch of computer vision,it plays an important role,which can effectively reduce computing resources and greatly improve the utilization of resources.At present,a large number of saliency detection models have been proposed by relevant researchers,but the application of the model is relatively scarce,especially for children with autism spectrum disorder attention exploration.On the one hand,due to the difficulty of collecting data,there is no large-scale experimental data.On the other hand,there is no visual saliency detection model for children with an autism spectrum disorder,which leads to the research progress has not been improved.Therefore,how to effectively detect the visual concerns of children with autism spectrum disorder has become one of the research hotspots,which opens up a new way for the treatment of autism.In addition,compared with the existing image saliency detection algorithms,there are some problems in the edge and location of the target,so how to detect and locate the target and its edge from the complex background is still a huge challenge.Therefore,aiming at the observed image data of autistic patients and the open image dataset,combined with deep learning technology,this paper proposes the corresponding algorithms of saliency detection based on a deep neural network.This paper mainly studies the following two aspects.(1)Firstly,aiming at the image data observed by autism spectrum disorder,this paper proposed a visual saliency prediction for autism spectrum disorder based on semantic features,and an end-to-end learning network model is proposed.Due to the small amount of image data,in order to avoid over fitting,this algorithm first enhances the image data,and then converts the original data by flipping,clipping and translating,etc.In addition,a loss function of positive and negative sample equilibrium is designed to achieve a better training effect.Through quantitative analysis and experimental results,the algorithm has good performance and high detection accuracy.(2)Secondly,for common image datasets,this paper proposed a salient object detection algorithm based on multi-label supervised learning.In order to better retain the edge information and positioning information of the target,a total of three sub modules are designed,which are a significant target monitoring module,edge information monitoring module and target positioning monitoring module,for detecting more accurate results.Finally,in order to further improve the performance of the model,the output saliency maps and feature maps are input into the refinement network for optimization,and the final stable saliency maps are obtained.By observing the experimental results on five common datasets,the model has good performance.