Saliency Analysis And Its Application In Object Detection

Salient object detection,aiming at modeling human visual attention mechanism,locates the most distinctive objects in an image.As one of the most fundamental task in the vision community,salient object detection and its applications in object detection have important value in many fields,such as public security prevention and control,network video and others.Deep learning based saliency detection algorithms learn from training data to obtain powerful feature representations,and accurately locate salient targets and overcome the impact of the complex background in an image.But deep learning based saliency detection also has some problems such as intra-class similarity and inter-class differences,low computational efficiency and rely on a lot of high quality training data.To adress these problems,this paper proposes a new network framework,structural loss function,and a saliency detection framework with weakly supervised learning.Beyond that,a tiny detection algorithm with a multi-level evaluation mechanism is proposed to improve the recall rate and computational efficiency.All algorithms are proved to be effective by experiments.To cope with the problems of insufficient feature representations and low computational efficiency,this paper proposes a network architecture based on hierarchical iterative feature fusion and multi-layer supervision.First,the residual block is used to imporve the network generalization;Then,the feature aggregation node is used to reduce the number of feature channels to improve the computational efficiency.Finally,each module of the framework outputs the prediction results,and the results are fused with a non-linear mapping method.To adress the problem of lacking supervision to the relationships of pixels in image,this paper proposes a label-based structure loss function.The relationship between image pixels is used as supervising information,inter-class supervision is used to highlight the differences between saliency target areas and background areas,and intra-class supervision is used to enhance the consistency of object and background pixel classification.To overcome the difficult of obtaining a large amount of high-quality training data,this paper proposes a weakly supervised saliency detection algorithm with the point label.Firstly,the superpixel algorithm is used to extend the point annotation to the region annotation.Then,we use the superpixel algorithm and fully connected conditional random field to optimize the training data,and using iterative training paradigm to improve the performance gradually.To relieve the difficulty of tiny detection caused by the lack of texture and detialed information,a saliency and compactness based tiny detection algorithm was proposed.Based on the compact characteristics,we cull low-quality proposals gradually with a cascade detection framework.First,we quickly obtain the objectness image with gradient.Base on the objectness image,we get all object proposals on the image.Secondly,using the minimum directional contrast(MDC)to obtain the saliency probability of pixels in proposals.Base on the saliency result of proposals,we measure the saliency probability distribution of the proposal.Finally,Base on the binary image dealed with OTSU,we adjust the boundaries of proposals to get better proposals.The evaluation method give the probability of the tiny object included in the proposals.Experiments show that the proposed method in our paper can achieve better detection rate,and has good computational efficiency.