Research Of Image Segmentation Based On Visual Attention Mechanisms

It has extensive application in many fields to automatically find objects of interest in images, such as compression, indexing and retrieval, re-targeting, and so on. There are two classes of such algorithms-those that find any object of interest with no prior knowledge, independent of the task, and those that find specific objects of interest known a priori. The former class of algorithms tries to detect objects in images that standout, i.e are salient, by virtue of being different from the rest of the image and consequently capture attention. The latter class of algorithms detects specific known objects of interest and often requires training using features extracted from known examples. In this thesis various aspects of finding objects of interest under the topics of saliency detection and object detection are addressed. The main content and the original techniques are as followed:Unified visual attention model evaluation was proposed to work out the existence of unfair problem on different data sets and different evaluation parameters. Synthetic images dataset, Bruce and Tsotsos-120dataset, HOU-62dataset and Achanta-1000dataset were chosen to test all kinds of algorithms under Precision, Recall, F-measure、Areas Under Average ROC Curves (AUC) and Linear Correlation Coefficient (LCC) scores. The experimental results show that, the contrast in the same standard was more fair, and different visual attention models could be distinguished more significantly. A visual attention model based on phase spectrum of color and intensity combined with threshold methods and morphology operator in the Lab color space was presented to segment salient objects. From scores such as Precision, Recall, F-measure, AUC and LCC, the performance of the model presented was better than the other models. The method was applied to image seam carving technique based on content, the experimental results show that, this method can better protect the salient target, background with large deformation, and keep the global and local image features.An improved method of distance regularized level set evolution based on visual attention mechanisms was proposed to segment objects. The method defined the focus points using ITTI model, and the object boundaries were acquired using distance regularized level set evolution. It could detect boundary of single object or boundaries of multi-objects and position accurately. And it also has strong anti-noise ability and could segment the images with weak boundary. From the results of experiments, the method was better than DRLSE and ACV under mean error.A visual attention model based on principal component analysis (PCA) was presented based on the characteristics of human visual attention mechanism with sparseness. Color and spatial features were extracted to combine the globe saliency map. The performance of the model was better than the other model under Precision, Recall, F-measure、AUC and LCC scores. Then, the defective on the solar cell was recognized based on a combination of PCA and radial basis function neural network (RBFNN). The testing samples recognition accuracy ratio of the PCA-RBFNN is96%on the dataset. The computation time was short to detect online.