Research On Image Saliency Detection Based On Deep Learning

Image saliency detection is one of the hot areas in computer vision,the main purpose of which is to extract regions of interest from images and locate salient areas.Image salient detection algorithm can be used as a preprocessing module for other computer vision algorithms,so it is of great significance to study the image significance algorithm in depth.In this paper,the image significance detection algorithm based on deep learning is studied,a multi-scale feature extraction module based on interactive learning is designed,and an F-measure like loss function and a loss function to enhance the consistency of image regions are given,which effectively improves the performance of image significance detection.When performing multiscale feature extraction,pooling results in a lack of texture detail.In order to solve this problem,this paper proposes a multi-scale feature extraction module based on the idea of interactive learning.Firstly,by learning the features of adjacent levels,the texture information in shallow features and the semantic information in deep features are fully integrated.Then,multi-scale information is extracted from single-level features to improve the robustness of significant area changes.Finally,in five datasets,the proposed module is experimentally verified to be superior to the current popular method.The pixel-level cross-entropy function cannot be perfectly adapted to the specific task needs of image saliency detection.To solve this problem,two loss functions are proposed as the objective function of image significance detection.Firstly,the non-derivable Fmeasure function is approximately compressed into a continuously derivable F-measure like function,which effectively prevents the gradient explosion problem and enhances the contrast of the overall significance map.Then,in order to enhance the consistency of the significance region during multi-scale feature fusion,the consistency enhancement function is given,which effectively copes with the problem of unbalanced before and after scenes in multi-scale features.Finally,the above two loss functions are balanced as the total loss functions of the third chapter model,and the effectiveness of the loss functions is verified by experiments.The experimental verification of the proposed method on the public dataset of image significance detection shows that the quantitative,qualitative and visualization results show that the module and loss function proposed in this paper are superior to other.