Research On Key Techniques Of Accurate Salient Object Detection In Complex Natural Scenes And Its Applications

With the popularization of products(e.g.,camcorders and infrared imaging devices)and technologies(e.g.,unmanned aerial vehicles and inspection robots),smart grids prefer to record and deliver information by using pictures,which provides convenience for modern society.Different from the human vision perceptual system,computers have difficulty in perceiving and understanding an enormous amount of visual information intelligently.Based on visual attention mechanism,people can catch attractive areas in an image quickly and effortlessly,while filtering out unimportant content.Salient object detection aims to help computers build up such an ability so as to make subsequent vision tasks focus on processing important image areas,which can avoid the superfluous computations of irrelevant visual information.The high robustness and accuracy of salient object detection algorithms are the most important precondition for serving as a pre-processing step for other vision tasks or more applications.For this purpose,developing an accurate saliency detection system which can be used in complex natural scenes has become one of hot research topics in computer vision.This dissertation makes intensive research on salient object detection by exploring propagation model and deep learning techniques.The goal is to design accurate and robust detection methods.Furthermore,considering that our country has made endeavors to promote the development of smart grids in recent years,many researchers are working on the electric power inspection by making use of the advanced computer vision technology.Thus this dissertation also explores the application of salient object detection in the power system.In details,the contents of this dissertation are as follows:1.To make up for the deficiency of most existing label propagation methods based on background prior,a method based on Page Rank and local spline regression is proposed for saliency detection.The input image is represented as two-scale graphs with homogeneous superpixels in order to explore the multi-scale saliency and saliency integration strategy.Multiple color features and spatial information are effectively captured to define the relevance of each node to its surroundings.Page Rank is used to assign the saliency value to each region depending on the similarity of the image elements with boundary cues.After selecting more robust background nodes and foreground nodes,a classifier based on LSR is employed to highlight more complete foreground regions.Extensive experiments on three large public datasets demonstrate that the proposed algorithm consistently achieves superior performance compared with most of unsupervised models.2.To overcome shortcomings of conventional approaches with handcrafted features,a novel multi-scale deep encoder-decoder network is proposed for salient object detection.Traditional approaches usually rely on handcrafted features,resulting in the lack of the semantic representation.To deal with this issue,the proposed algorithm adopts deep convolutional neural networks to model visual saliency.However,early CNN-based models fail to precisely segment the boundaries of salient objects from a cluttered background due to the downsampling effects or the patch-level operation.To handle this,the proposed model uses an end-to-end way to generate fine-grained results from a coarse to fine level.To guide the deep model in learning more discriminative features,a semantic-guided module is designed to avoid the dilution of informative features from the encoder network.Moreover,a multi-scale mechanism is utilized to help deep saliency model extract multi-scale features.Extensive experiments on six challenging datasets demonstrate that the proposed model outperforms eleven mainstream methods.3.In the purpose of improving the feature extraction ability of deep convolutional neural networks,a novel context-aware attention network is proposed for salient object detection.By fusing local and global contexts,the designed network integrates context-aware attention modules that detect salient objects by simultaneously constructing connections between each image pixel and its local and global contextual pixels.Specifically,each pixel and its neighbors bidirectionally exchange semantic information by computing their correlation coefficients,and this process aggregates contextual attention features Both local and global contexts are exploited to learn complementary contextual features that help filter out responses from noisy backgrounds and highlight salient foreground objects effectively.Extensive experiments show that the proposed model demonstrates superior SOD performance against most of the current state-of-the-art models.4.The proposed CNN-based saliency models are applied in several electrical fields,including insulator segmentation,free space detection and smoke detection.First,sufficient number of images collected from specific electrical scenario constitute a complete dataset.Second,the pixel-level annotation is conducted on the electrical dataset.Third,the deep saliency models are retrained on the training images and evaluated on the testing images.Extensive experiments show that salient object detection can be well adapted for electric power vision tasks,which has potential application prospect in the future.