视觉重构脑解码理论与方法研究

发布时间：2018-03-04 06:24

本文选题：视网膜拓扑映射　切入点：视觉图像重构　出处：《电子科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：视觉系统是人类获取外部环境信息的重要感官之一,人脑视觉信息处理具有高效、抗噪和鲁棒性,尤其表现在对自然场景的识别和处理上。大脑惊人的视觉信息处理能力背后,隐藏着怎样的神经机制呢?这是令无数科学家着迷的问题,解读人脑视觉认知快速编码/解码机制一直是脑科学最具前沿和挑战性的方向。随着脑编码机制的逐渐清晰和脑成像技术及模式识别方法的快速发展,科学家们开始思考更复杂更深层次的问题—脑解码:能否根据脑功能活动信号进行视觉重构,推算出大脑所知觉到的视觉信息,比如文字和图像?从2005年开始,国际上一些著名科学家和实验室开始探索脑解码的方法和技术,近年来取得了很大进步。Miyawaki利用早期视皮层视网膜拓扑映射关系建立了一种不受限于候选类别的脑信号视觉图像重建方法和技术,但是,他的方法重构出的图像具有很大的噪声,并且重构的效率低。针对这样的不足之处,本文提出了基于F-score特征选择的贝叶斯重构方法,来改善图像重构的精度和效率。要实现视觉图像的重构,则需要对大脑的视觉皮层亚区的定位,本文首先开展了视网膜拓扑映射的磁共振实验。该实验通过“楔形”和“环形”刺激来诱发大脑视觉皮层,使用广义线性模型(GLM)来找到视觉皮层中激活体素,计算被激活体素的极角值和离心率值,得到大脑的极角图和离心率图。根据极角图和离心率图初步实现了初级视觉皮层亚区的划分,实现了V1,V2,V3区域的体素定位。针对Miyawaki提出的基于SMLR图像重构方法噪声大和重构效率低的问题,本文建立了一个基于F-score特征选择算法的贝叶斯视觉图像重构模型。该模型使用F-score特征选择方法选择出重构效果好的体素,去除了与刺激图像不相关的体素,最终降低了重构图像的噪声。结合复杂度低的贝叶斯方法,能显著提高视觉图像重构的效率。
[Abstract]:Visual system is one of the important senses for human to obtain external environmental information. Human brain visual information processing has high efficiency, anti-noise and robustness, especially in the recognition and processing of natural scenes. What kind of neural mechanism is hidden? It's a fascinating question for countless scientists, It has been the most advanced and challenging direction of brain science to interpret the mechanism of fast coding / decoding of human brain visual cognition. With the development of brain coding mechanism and the rapid development of brain imaging technology and pattern recognition methods, Scientists are starting to think about a more complex and deeper problem-brain decoding: can we reconstruct the brain from the signals of brain activity and extrapolate the visual information the brain perceives, such as words and images? Since 2005, some famous scientists and laboratories around the world have begun to explore the methods and techniques of brain decoding. In recent years, great progress has been made. Miyawaki has established a new method and technique of brain signal visual image reconstruction, which is not limited to candidate categories by using the early retinal topologic mapping of visual cortex. However, the reconstructed image by his method has a lot of noise. And the efficiency of reconstruction is low. In this paper, a Bayesian reconstruction method based on F-score feature selection is proposed to improve the accuracy and efficiency of image reconstruction. Then we need to locate the subregion of the visual cortex of the brain. In this paper, we first carried out a magnetic resonance imaging experiment of retinal topologic mapping, which induces the visual cortex of the brain through "wedge" and "ring" stimuli. Using the Generalized Linear Model (GLM) to find activators in the visual cortex and calculate the extreme angle and centrifugation rate of the activated voxels. The polar angle map and the centrifuge rate map of the brain were obtained. According to the polar angle map and the centrifugal rate map, the primary visual cortex subregion was preliminarily divided. The voxel location of V1 / V2 / V3 region is realized. Aiming at the problems of Miyawaki's image reconstruction method based on SMLR, which has high noise and low reconstruction efficiency, the voxel location of V1 / V2 / V3 region is realized. In this paper, a Bayesian visual image reconstruction model based on F-score feature selection algorithm is established. Finally, the noise of reconstructed image is reduced, and the efficiency of visual image reconstruction can be improved significantly by combining the Bayesian method with low complexity.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;R338

【相似文献】