Visual saliency detection method combining machine learning, background suppression and perception positive feedback

Abstract

The invention discloses a visual saliency detection method combining machine learning, background suppression and perception positive feedback. An algorithm framework is put forward by simulating human eye microsaccade and perception recession mechanisms. An image is directly and roughly divided into a gazing region and a non-gazing region; and multi-times random sampling is carried out on pixels in the two regions to simulate repeated scanning of the gazing region by microsaccade. For a plurality of sample sets after sampling, a plurality of PELM models are constructed by learning; and classification results of multiple models are superposed to form a rough saliency graph. Background suppression is carried out on the rough saliency graph by using an RBD algorithm and a positive feedback iteration process based on PELM is constructed for the gazing region; and if the PELM classification result in iteration is stable, the visual sensing effect is saturated and circulation is ended. The PELM classification result can be viewed as visual stimulation and after stimulation superposition, a new saliency graph with a target enhanced is formed. Therefore, step-by-step-refinement saliency detection driven completely by data can be realized.

Description

technical field [0001] The invention relates to the technical field of human visual simulation, and specifically uses machine learning for real-time online modeling to construct a fully data-driven visual salience automatic detection method. Background technique [0002] Visual salience is the basis of visual attention. Visual saliency detection is a hot issue in the field of computer vision research today. How to simulate the human brain-eye system and realize machine vision that simulates human vision has always been the research direction of researchers. Natural images are typical unstructured data, and machine learning is suitable for modeling unstructured data. In recent years, machine learning algorithms constructed from shallow and deep neural networks, based on bottom-up and top-down frameworks, have been applied to solve problems such as visual saliency detection. The bottom-up framework can adopt a data-driven approach to modeling, but the algorithm is usually c...

AI Technical Summary

Problems solved by technology

The bottom-up framework can adopt a data-driven approach to modeling, but the algorithm is usually limited by some prior knowledge, and the model is prone to deviation due to the often inaccurate training samples

In the top-down framework, deep learning has been successfully used for image segmentation and saliency detection; so far, deep learning-based algorithms have reflected the best performance in many applications, but their shortcomings are also obvious: deep learning A large amount of labeled sample data is required, and deep networks often need to be manually designed. Compared with shallow neural networks, their training requires higher hardware resources, which cannot be performed online in real time at present

[0003] Clearly, existing methods may fail to implement without suitable prior knowledge and sufficient valid samples in practice

In addition, we noticed that in most current saliency detection methods, the information is often one-way mapped, lacking a dynamic feedback process, which is very different from the human visual system, which may be the reason why the current machine vision is far from human vision one

Images