Research On Salient Object Detection Algorithm Based On Deep Learning In Complex Scenes

Salient object detection extracts the most visually salient object or area in the image.It is an important content in computer vision and related research fields,and it can be applied to tasks such as object tracking,traffic control and image retrieval.The salient object detection algorithms based on traditional handcraft features can be effectively applied to general scenes.Due to the lack of sufficient semantic information,these algorithms are no longer applicable in more complex scenes.Therefore,salient object detection in complex scenes still needs further research.With the development of deep neural networks,researchers have applied deep semantic features to salient object detection.However,existing algorithms based on deep learning still have challenging issues in complex scenes.For example,the area which has salient appearance in complex backgrounds is easily identified as a salient object.Therefore,aiming at detecting salient objects in complex scenes,the thesis focuses on extracting and using the deep semantic features that can effectively distinguish salient objects to improve the performance of salient object detection algorithms.The main research work of the thesis is as follows:1.A salient map refinement algorithm based on global pixel features(SMR_GPF)is proposed.The algorithm designs a full convolutional neural network based on the VGG16 network,and the initial salient map obtained by super-pixel segmentation is combined with the deeper feature map of the network in the form of a single channel feature map.After connecting,the new feature map continues to propagate forward through the convolution network.Finally,the softmax function is used to distinguish the pixel-level salient regions.When training the model,the back-propagation algorithm updates the network parameters before and after the feature connection from the back to the front in an end-to-end fashion.The pixel-level marking of the algorithm avoids the problem of inconsistent markers within the super-pixel and enhances the performance of the model in detecting salient regions by learning the depth semantic features of the image.Verification is performed on five image datasets containing multiple salient objects and complex backgrounds.The experimental results show that the proposed algorithm can effectively reduce the detection error of salient objects in complex scenes,especially for the test images with complex background.2.A salient object detection algorithm based on multi-feature fusion(SOD_MFF)is proposed.The algorithm first trains a salient nomination network,which extracts multiple candidate regions in the image that contain salient objects.By using the spatial information and category confidence value of the candidate region,a mixed Gaussian distribution map is generated as a position prior distribumap of the network,the position prior distribution of the salient object and the initial salient tion of the salient object.The prior distribution is embedded into the deep neural network as a single channel feature map.At the deeper feature map are combined to obtain a new feature map.The proposed model finally use the forward propagation algorithm to predict a pixel-level salient object or region.When training the model,the whole network as an end-to-end structure simultaneously updates all parameters of each layer to obtain optimal results.The algorithm embeds a priori distribution of salient regions in the network.The prior distribution separates the complex background region from the salient region,which reduces the learning difficulty of the neural network.Finally,the use of depth semantic features improves the performance of the model in detecting salient regions.Verification is performed on four image datasets with multiple salient objects and complex backgrounds.Experimental results show that the algorithm effectively improves the detection accuracy of salient objects in complex scenes,especially for the test images with complex backgrounds.