Research On Salient Object Detection Technology Based On Deep Neural Networks

Salient object detection is an important research topic in the field of computer vision.It aims to detect and segment the most salient region(s)or object(s)of an image by analyzing its content.Other than being a standalone task,salient object detection can also be used as a pre-processing algorithm to improve the performance of other computer vision tasks.Therefore,it is of great importance to conduct an in-depth research on salient object detection.At present,convolutional neural network has achieved great success in the field of image processing due to its strong expressive ability for nonlinearity.However,the following problems still exist: 1)The current algorithms are insufficient dealing with contours of objects in complex scenes.2)The extraction and fusion of salient features are not sufficient.3)The current research on visual saliency is limited to salient object detection without further exploration into related topics(e.g.saliency ordering).This dissertation carries out in-depth research on visual saliency detection,and the main content and innovative points are summarized as follows:1.A salient object detection method based on contour self-compensated network is proposed,namely CSCNet.It can compensate the network profile information from three aspects.Firstly,the super-pixel map is used as an auxiliary input to strengthen the contour information.Secondly,the salient object detection problem is reconstructed as a multi-classification problem of the background,objects and contours,in which the contours of the salient objects are used as the third label to mark the data.Thirdly,a contour penalty loss is proposed based on the spatial relationship of three types of labels for supervising the training of the network.The proposed CSCNet is evaluated on seven standard salient object detection datasets.Compared with other deep salient object detection methods,CSCNet can ensure the integrity of the salient object contour more effectively without adding too many convolutional layers and parameters.2.An attentive feature-based feedback network is proposed for salient object detection in this dissertation,namely AFFNet.A multi-scale feature fusion module(MFF) and a channel-wise attention module are used to obtain and fuse the low-level and high-level feature map separately.A feature-wise attention module is used to map the features into the feature extraction network to enhance the robust of extractor,in which way,the feedback network is built.The proposed AFFNet is evaluated on 6 standard salient object detection datasets.Compared with other state-of-the-art deep salient object detection methods,the proposed method has excellent generalization ability and better robustness for predicting salient object.3.Driven by business intelligence applications for rating attraction of products in shops,a new problem--salient object grading is studied in this dissertation.On this basis,a pixel-wise ordinal classification method is proposed,namely POC.It consists of a multi-resolution saliency detector,which detects and segments objects,an ordinal classifier,which grades pixels into different salient levels,and a binary saliency enhancer,which sharpens the difference between non-saliency and all other salient levels.Two new image datasets with salient level labels are constructed,namely SALICON-OR and ILSO-OR respectively.Experimental results demonstrate that,on the one hand,the proposed POC method can not only effectively estimate the saliency rank of objects in the image,but also provide the saliency grading of the objects between the images.On the other hand,the grading information of the POC algorithm helps to improve the binary salient object detection in the traditional saliency detection problem,and offers very comparable performance with state-of-the-art salient object detection methods.