Salient Object Detection And Application Based On Attention Mechanism Convolutional Neural Network

With the popularization and development of smart tools such as mobile phones and computers,digital image information plays a great role in everyday communications.According to statistics,billions of images are transmitted to the network every day,which carry out the life of people.Therefore,we need to strengthen the research on computer vision,the computer can help people to selectively process the image information in the network.The important object region is retained in image,and the backgound information that is interfernce with information is eliminated in image.It is defined as salient object detection.Since the emergence of deep learning,salient object detection has developed rapidly,but it still cannot fully meet the standards of human visual recognition.Therefore,this thesis continues to conduct in-depth research on salient object detection in computer vision.The main work of this thesis is as follows:(1)Salient object detection based on attention mechanism convolutional neural network(AMCNN)In this thesis,attention mechanism is introduced into convolutional neural network for salient object detection.Firstly,the channel attention module is added into the deep features of NLDF network.The global average pooling layer of the channel attention module control to intemperately weaken background semantic information.Two full connection layers implement the salience mapping of deep semantic information.The object information with more complete content is obtained.Secondly,spatial attention module is used to the extracted features of the channel attention module.The convolution operation in spatial attention module extract the in-depth features of the image.The spatial information of low-level features which restrict the semantic information of deep features is obtained by multiplying the extracted results with low-level features.It is realized the integrity optimization of the content of significant object area.Finally,the bilinear interpolation method is employed the last layer of the in-depth features in NLDF network.which replaces the original global feature in 1×1 scale.It can preserve more detail information to entirely optimize the content of the object region,It solves the problem that NLDF method lacks the content integrity of the object region,that is,the complete object information cannot be obtained in some complex scenes,especially the missing or weakening of the information which is at the edge of the object.Compared with 10 classical salient object detection algorithms in recent years,the experimental results show that the salient object detection algorithm based on AMCNN(Attention Mechanism Convolutional Neural Network)proposed in thesis can perform the same or better performance as the advanced algorithm without loss of detection speed.(2)Engineering application of image retrieval based on salient object detectionIn order to further verify the value of salient object detection's engineering application,AMCNN salient object detection method is applied to image retrieval engineering in this thesis.AMCNN salient object detection model is added to the preprocessing step of image retrieval project.For any retrieval image,it needs to extract their features through the pretreatment of AMCNN firstly,and select the important information of retrieve image based on the position of the object area,reduce the useless background information,so as to increase the description of image contents' correlation,achieving more efficient image retrieval engineering application.