Fitting-based Optimization For Image Visual Salient Object Detection

Image visual salient object detection plays a major role in many image processing and computer vision applications,such as image compression,classification,and retrieval.Researchers have introduced various salient object detection algorithms.However,saliency maps generated by each existing salient object detection algorithm usually show similar imperfection against the ground truth.By gradually improving the performance of optimization algorithms,four optimization algorithms for salient object detection algorithms are proposed in this thesis.And the performances of our saliency optimization algorithms are evaluated using traditional evaluation metrics,image quality assessment metrics,and the content-based image retrieval application.Firstly,we propose a statistical fitting-based optimization method for salient object detection algorithms.The algorithm first computes the statistics of the ground truth and saliency maps computed by each salient object detection algorithm.We then use those statistics to compute the parameters of fitting models,which generally agree with statistical data characteristics.For a new saliency map,we use the fitting model with the solved parameters to obtain the fitted saliency values.Experimental results show that the proposed algorithm is suitable for optimizing a variety of salient object detection algorithms and the quality of the optimized saliency maps improves in terms of traditional evaluation metrics,image quality assessment metrics,and the content-based image retrieval application.Secondly,we present a clustering and fitting-based optimization algorithm for salient object detection by using the similarities among images.The algorithm uses the K-means method to cluster the images into k clusters according to the similarities among images.Image similarity is measured in terms of scene and color by using the GIST and color histogram features.The saliency map of a new image is optimized by using one of the fitting models which corresponds to the cluster to which the image belongs.Experimental results show that our clustering and fitting-based optimization algorithm improves the performance of a variety of salient object detection algorithms.And the improvement achieved by algorithm when using both clustering and fitting strategies is greater than the improvement achieved by the same algorithm when not using the clustering strategy.Furthermore,this thesis proposes a neighbours learning-based saliency optimization method for salient object detection by using the similarities among neighbouring pixels.The algorithm exploits machine learning method to model the relationship between ground truth and saliency maps and among saliency values of neighbouring pixels.And we apply the fitting model to a new saliency map to obtain the optimizated saliency map.The experimental results illustrate that the proposed algorithm is suitable for optimizing many salient object detection algorithms.And the performance of the optimization algorithm is further improved.Finally,in order to optimize the saliency maps of images with multiple objects in natural environment,we present a conditional random field-based optimization algorithm for salient object detection.The algorithm uses fully connected conditional random field method.It combines the color and position information to adjust the original value of the saliency map,and eliminates the impact of noise.The image clustering and grid search methods are used to search the best weights of kernel functions in conditional random field,and the saliency map in the same cluster is optimized using the same parameter settings.Experimental results show that the conditional random field based salient object detection optimization algorithm is not only suitable for optimizing the saliency maps for image in simple scene datasets,but also for optimizing those in complex scene datasets.