Research On Salient Object Detection Of Nighttime Scenes

Image saliency detection aims at simulating the human visual mechanism,which can extract the most attractive objects or regions in the image.This research has provided new ideas for the rapid and effective processing of visual data and has become a hot topic in the field of computer vision.At present,many salient object detection models were proposed at home and abroad,and most of them are focus on the visible light scene,they seldom consider the nighttime scenes.Due to low lightness,degraded color information,low contrast,and low signal-to-noise ratio,the image perception quality is greatly reduced.Therefore,existing algorithms and techniques have great difficulty in analyzing,understanding and detecting the nighttime images.Thus,the detection of salient object in the nighttime scenes becomes a challenging research,and the solution of this problem can provide theoretical and technical basis for hot issues such as nighttime security monitoring and object targeting in complex environment.Based on the existing salient object detection research,this thesis analyzes the deficiencies of these models,and discusses the main attributes of the nighttime scenes with visual attention mechanism.Then,several visual saliency models for nighttime scenes are proposed.The nighttime image dataset is constructed to verify the effectiveness of the proposed algorithms.The main contents and innovative points are as follows:(1)Aiming at the problem that the saliency map generated by nighttime image in a single space is incomplete,a discrete stationary wavelet transform(DSWT)based saliency information fusion model is proposed in this thesis to integrate the frequency and spatial domain information from the pixel-level.The proposed algorithm mainly consists of three steps,of which the frequency domain based saliency measure,the spatial domain based saliency measure and the DSWT based non-linear fusion are considered.By analyzing the sensitivity of human visual system to the light of different frequencies under low light conditions,the frequency and spatial saliency maps are calculated based on the HSV color space which has high perceived intensity in nighttime scenes,then DSWT is used to combine the redundancy and complementary frequency domain and spatial information in a non-linear pixel-level fusion way.Thisframework is simple,effective and easy to extend,and it realizes the initial detection of salient objects in nighttime scenes.(2)Aiming at the problem that the discrimination performance of the feature extracted from nighttime scenes has great difference,a superpixel covariance salient object detection model is proposed to fuse the optimal hand-crafted features in nighttime scene from the feature-level.The proposed algorithm is mainly composed of five parts,namely,the construction of superpixel based graph model,manifold ranking,regional covariance based feature fusion,local-global saliency measure,and global search based saliency optimization.Since the regional covariance has strong adaptability to rotation,scale scaling and illumination changes,it effectively preserves the feature information that is less sensitive to nighttime scenes during the nonlinear fusion,thereby it greatly improving the accuracy of salient object detection.This model overcomes the problem of feature failure in nighttime scenes by considering the non-linear fusion of feature levels.It can not only generate saliency map with clear boundary and full resolution,but also effectively restrain the background interference and enhance the performance of salient object detection in nighttime scenes.(3)Aiming at the problem that the training information is unbalanced due to the single visual information in the nighttime image,the deep feature is introduced based on the hand-crafted feature,and a region covariance guided convolution neural network(CNN)model is proposed in nighttime scenes from the feature-level.The proposed algorithm mainly contains four parts: the extraction of multiple features,the construction of covariance matrix,CNN-based sample training and contrast-based saliency calculation.The covariance matrix of hand-crafted low level feature is used as the input sample for deep learning,thus the correlation between the samples is trained,and the complex features are learned to express the saliency information of the nighttime image block.The low recognition and background interference in nighttime scenes can be effectively overcomed by the strong expression ability of the deep framework,and the accuracy of object detection is further improved.(4)Aiming at the problem that the structural information and the boundary information of obtained salient object by deep model cannot be effectively retained in nighttime scenes,a global convolution and boundary refinement guided deep convolutional neural network is proposed to detect the salient object from thedecision-level level,the proposed algorithm is mainly based on the deep features and is composed of three parts,a deep full convolution structure based on local-global features,a global convolution module and a boundary refinement module.After random initialization,more abundant local and global features from a large number of image data can be automatically learned by the full convolution network.The global convolution module enables the network to obtain more precise object structure features,thus the discriminant performance of the model is enhanced.The boundary refinement module enables the model to retain more complete boundary information.The proposed model has stronger discrimination and generalization performance in nighttime scenes,while the detected salient objects have a more uniform internal structure and a more complete edge information.To sum up,through the above research,this thesis constructs a nighttime image dataset,and puts forward four kinds of saliency object detection algorithms for nighttime scenes from the pixel-level,feature-level and decision-level.It has important reference value for the research in the field of computer vision.