Saliency detection method based on visual perception positive feedback

Abstract

The invention discloses a saliency detection method based on visual perception positive feedback, comprising: 1) using various existing saliency detection methods to detect image saliency initially; 2) superposing the above results to generate a new comprehensive saliency image, and binarizing the image by means of thresholding method to form binary fixation area Ip; 3) acquiring few pixel samples inside and outside the Ip repeatedly, training, and parallelly constructing various RVFL (random vector functional link) neural network models; classifying pixels by the multiple neural network models, and integrating to form binary target output BW; 4) returning the BW as a neural fired pulse to step 2, and superposing with the comprehensive saliency diagram to form an iterative cycle; 5) during iterating, if the input Ip and output BW in positive feedback link are substantially the same, indicating that perception is saturated, and stopping iteration. Ip or BW is the most salient target segmentation result in the image. Human vision is simulated by superposition of various saliency detection methods and iteration of visual perception positive feedback, and a visual saliency image more like human visual perception is obtained.

Description

technical field [0001] The invention relates to the technical field of human visual simulation, in particular to a saliency detection method based on positive feedback of visual perception. Background technique [0002] Traditional image processing algorithms suffer from problems such as ever-changing scenes, massive data, and high-dimensional features, and have obvious limitations. The performance of the human visual system far exceeds the current algorithm, and simulating the principle of human vision is an effective way to break through the current algorithmic dilemma. Human vision has evolved over a long period of time and has an active vision mechanism that focuses on objects of interest in the scene through visual attention. Visual attention models are the starting point for researchers to simulate human vision, and can be divided into two types: data-driven and task-driven visual attention models. [0003] Data-driven models perform bottom-up attention, computing sa...

AI Technical Summary

Problems solved by technology

The disadvantage is that the deep learning network requires large-scale labeled training samples, the network structure is manually designed, the network performance depends on the training samples, the network training time is long, the requirements for computer hardware are high, and online real-time training is difficult

[0004] We noticed that in the existing visual attention models, the algorithm process usually lacks a dynamic feedback link, which is quite different from the process of human visual perception.

Images