Research On Salient Region Detection Algorithm In Natural Scenes

Image salient region detection is a popular research topic in the field of computer vision,simulate the human visual attention mechanism and select the most attractive region from the scenes effectively,thus significantly saving computer resources.As an important step of image preprocessing,salient region detection algorithms have been widely used in image compression,object detection,image retrieval,semantic segmentation and many other fields.With the development of deep learning,salient region detection technology has made great progress.However,natural scenes are very complicated,and existing detection algorithms still have various shortcomings.Based on the analysis and review of existing salient region detection algorithms,this thesis proposes two salient region detection algorithms for better performance.(1)The pooling layer makes the high-level output feature map of the convolutional neural network is smaller than the input,which is not conducive to the salient region detection tasks obviously.To solve this problem,this thesis adopts Deep Lab V3+ as the basic network,and proposes a salient region detection algorithm based on encoder-decoder network.The encoder and decoder is used to extract high-level semantic information and restore the size of the feature map,respectively.In detail,the decoder is firstly proposed to make full use of the low-level information and recover the size of the feature map gradually,which overcomes feature loss after the pooling layers effectively.Then,during the training process,the binary classification cross-entropy loss function is adopted during training process and preprocesses the input data to highlight the differences among individuals.By this way,the algorithm becomes more suitable for salient region detection task.Finally,comparison with other algorithms on six public datasets shows that the proposed algorithm has better performance.(2)Aiming to the rough object contour by salient region detectors in complex natural scenes,a novel algorithm is proposed.In essence,it is a deeply supervised salient region detection algorithm that combining object contour features.The algorithm performs salient region detection and contour recognition task simultaneously,and applies the contour features to optimize salient object features gradually.First,it adopts side output sub-networks to extract multi-scale features from VGG16 and predicts them to obtain salient object features.Then, a contour recognition network with deep supervision is used to obtain contour features that match salient target features.Finally,the salient object features are gradually refined by object contour features.The final results can be obtained by fusing the saliency maps of each layer according to different weights,which further improves the quality of the results.Experiments show that the proposed algorithm can effectively overcome the problem of rough object contour.