Saliency Detection In Images With Complex Background By End-to-End Sparse Maxout CNN

Describing objects or areas in images that grab human's attention is a simple and natural process for human being.However,it is a challenging task for computers because not only does it need to separate objects,people,or actions containing in images but also detect them in images with a complex background.In recent years,models with end-to-end architecture in deep learning have made great progress in saliency detection and prediction tasks.The main research work of this thesis as follows.An in-depth analysis of existing classical and novel salience detection methods was carried out.As a result,the shortcomings have motivated the development of a method based on endto-end sparse maxout convolution neural network for saliency detection in images with complex background.It also describes main key technologies that are used to solve the problem.Specifically speaking,these technologies include common neural networks,convolutional neural networks,and their regularization methods.In the process of designing and creating of the end-to-end sparse maxout convolutional neural network for saliency detection in images with a complex background,the classic maxout method was considered,and its main drawbacks were also described.Furthermore,the principle of the operation of the proposed method is described in detail,and the main advantages compared with the classic maxout version are highlighted.The effectiveness of the method has been experimentally demonstrated by creating two networks for processing MNIST and CIFAR10 datasets.As a result,the sparse maxout is more effective than the maxout,moreover,the proposed method retains the main advantages of the classical algorithm.In addition,the architecture of the proposed end-to-end sparse maxout CNN for saliency detection in images with complex background was demonstrated.This thesis proposes a saliency detection model for images with complex background based on the end-to-end sparse maxout convolutional neural network.We introduce the saliency detection model based on two main ideas.The first one is considering the sparsity of the convolutional neural network.The second one is implementing the end-to-end architecture for saliency detection.Experiment results on CAT2000,SALICON,MIT300,and iSUN demonstrate that the proposed method overcome state-of-the-art results in saliency detection and prediction tasks.Furthermore,the analysis of different saliency evaluation metrics related to the results of the experiment is provided.