Complex scene bottom-layer visual information extraction method

Abstract

The invention belongs to the field of scene visual cognition, and particularly relates to a complex scene bottom-layer visual information extraction method. The problem of bottom-layer feature value extraction of a complex scene is solved. According to the method, an improved convolutional neural network structure is introduced, and four convolutional filters are utilized to form a multi-depth analysis set to perform image semantic segmentation on a scene image. A feature convolution filter is introduced to screen and extract regional semantics of a complex scene, and the semantic region segmentation is performed on a scene image by using transposed convolution on an extraction result. The region segmentation result of the scene semantics is substituted into the final bottom visual information feature value extraction network as an activation bias, so that various types of scene details can be ensured not to be lost. After semantic segmentation of the scene area, the scene bottom-layerinformation feature value required by the cognitive experiment can be well extracted, and details in a complex scene can be well reserved.

Description

technical field [0001] The invention belongs to the field of scene visual cognition, and in particular relates to a method for extracting underlying visual information of complex scenes. The present invention is mainly aimed at complex visual cognition experimental scenes through algorithmic realization, and extracts three types of dominant underlying visual attention resources in the global scene. Background technique [0002] Human's underlying visual processing mechanism and neuroscience research on visual cells point out that when people observe the scene without the influence of existing concepts, human beings tend to allocate visual attention resources to areas with high color saturation in the scene, Areas of high color contrast and areas with edge / directional features. These three types of attention resource content are called overt underlying visual features. In the neural signals of human visual attention, these three types of features occupy most of the informati...

AI Technical Summary

Problems solved by technology

[0005] In view of the defect that the existing low-level visual information eigenvalue method loses the details of the eigenvalue during the extraction, this defect will cause the visual cognition experiment of complex scenes to fail to obtain effective low-level visual eigenvalues ​​during scene analysis, and the lack of this part will be extremely serious. The Reliability of High Impact Cognitive Analysis

Images