Methods Of Object-of-Interest Extraction Based On Biological Visual Attention

According to cognitive psychology, image area which can generate novel, stronger and expected stimulus is easy to draw the observers’attention. Therefore the study of obj ect-of-interest extraction inspired by visual attention mechanism has significant values in both theory and applications. This thesis mainly studies the methods of obj ect-of-interest extraction in static images based on sparse coding theory and biological visual attention mechanism. The main work of this thesis is as follows:1. The dictionary learning based sparse coding method is studied. Sparse coding is mainly based on dictionary learning, and the typical dictionary learning algorithm is the generalizing K-Means clustering algorithm based on singular value decomposition (K-SVD). However, K-SVD algorithm is time-consuming, therefore a fast dictionary learning algorithm is proposed, which is simple and fast, and can significantly accelerate the solution to dictionary learning.2. A visual attention model based on rarity feature is put forward for the problems existed in traditional saliency-based object extraction methods. Dictionary learning is embedded into the feature extraction process and the rarity features can be extracted. The proposed model can solve the problems that the target size is uncertain and the outline is fuzzy much better.3. An improved method of obj ect-of-interest extraction is proposed by introducing the sparse coding to ITTI model. This method deals with the image by utilizing the sparse coding first, then uses ITTI model to extract the object. Sparse coding can extract the rarity features of the target areas effectively, distinguish object areas from background areas accurately, and widen the pixel values between the two areas obviously. Experimental results show that the improved ITTI model can detect the target areas with low saliency very well.