Research And Application Of Salient Object Detection Technology Based On Deep Learning

With the development of information technology,image data as the main content of information dissemination is infiltrated into people's daily life.In order to more effectively and quickly obtain map that meets people's focus,researchers refer to the human attention mechanism and propose large number of image saliency detection algorithms that achieve significant results.But there are edge blurs and the model occupy a lot of storage resources,etc.For this reason,this paper studies the saliency detection of multi-scale and multi-level feature optimization.The research work is as follows:First,the saliency detection optimization based on multi-scale features.According to the investigates and analyzes of saliency detection algorithms proposed in recent years,the paper finds that existing saliency detection algorithms cannot consider both objection positioning and detailed description at mean time.In response to the above problems,based on the objection classification neural network VGG16,features have been strengthened on multiple scales,both shallow and deep features have been used to generate detailed optimization weights and objection positioning weights,respectively.Two weights are used to optimize the enhanced feature twice.Finally,a prediction saliency map is generated.In addition,the results of saliency detection optimization based on multi-scale features and other six outstanding methods made a comparison of three different datasets about PR curves,F-measures and MAE.Secondly,the saliency detection optimization model based on multi-level features.The purpose of saliency detection is to detect eye-attracting areas from images or photos.With the efforts of researchers,the saliency detection has made significant progress in comprehensive performance.But most of the existing models are bothered by the large storage space and slow detection speed.To deal with these problems,this paper proposes a model based on multi-level features optimization for salient object detection.The model uses 64-dimensional embedding features to simplify and learn multi-level image information at different stages of VVG16,while benefit from being more representative 64-dimensional embedding features.The model storage space is small(81MB)and the detection speed is fast(25fps).Finally,the proposed saliency detection model is applied to image compression.According to the investigation and analysis of existing image compression technology,the multi-level features optimization of saliency detection technology proposed in this paper is introduced into the image compression.Through compression and reconstruction experiments of multiple real scene images,it is proved that with the aid of saliency detection technology and optimized level features,high-quality compression of important area information in the image can be achieved.