毕业生知识库

视觉作为人类的一种重要感知模态，是人类获取信息的主要途径之一，人脑中大约70%的信息是来自于眼睛，大概20%-30%的皮层区域是用来做视觉处理。在长期的进化过程中，人的视觉系统已经成为功能完备、机制完美的信息处理系统之一，隐含了许多极为高效的视觉图像处理机制。

深度学习作为一种“粗糙”的类脑计算模型，它的学习过程需要大量的训练样本，而人类的视觉学习过程是基于小样本的，这充分体现了人类视觉的灵活性与高效性；此外，相比计算机视觉，人类的视觉处理过程是非常快速和鲁棒的。因此，研究视觉信息的加工处理过程，一方面有助于理解大脑皮层的工作原理，另一方面有助于提升现有计算机视觉的效率和性能。

神经解剖学和神经影像学的发展为模拟视觉皮层结构与机制，构建视觉系统的计算模型提供了丰富的生理依据。本文充分调研了初级视皮层和腹侧通路的结构功能特性，研究了基于腹侧通路的视觉计算模型，主要工作是：

首先，本文基于初级视皮层生物特性的MCI模型，融入了一些心理学现象和初级视皮层的生物机制，提出更加类脑的sMCI模型，以及sCMCI模型，在提升速度的同时保证了准确率。基于该模型，本文在自有的无人机数据和一些公开数据集上进行了验证，实验结果表明本文提出的方法在提升计算效率的同时保证了轮廓提取的准确性。

其次，本文研究了大脑皮层机理的分层时序记忆模型（Hierarchical Temporal Memory,HTM）,并基于HTM模型进行腹侧通路计算模型的构建，使其更符合视皮层的生物学特性，本文将其应用于手写体和“幻视”图片识别的任务中，取得了较好的效果。同时，HTM在“幻视”图片识别任务中的结果比卷积神经网络LeNet-5的结果更好，这也间接表明了HTM更符合人脑的视觉机制。

 

Vision, as one of the most important perceptual modes of human beings, is one of the main ways for human to obtain information. About 70% of the information in human brain comes from the eyes, and probably 20%-30% cortex area is used for visual processing. In the long process of evolution, human visual system has become one of the information processing systems with complete function and perfect mechanism, which implies many highly efficient visual image processing mechanisms.

As a "rough" brain-like model of computation, deep learning requires a large number of training samples. The human visual learning process is based on small-sample, which fully embodies the flexibility and efficiency of human vision. In addition, compared with the computer vision, the human visual processing process is very fast and robust. Therefore, the research on the processing mechanism of visual information is not only helpful to understand the working mechanisms of the brain, but to improve the efficiency and performance of the computer vision.

The development of neural anatomy and neuroimaging provides a rich physiological basis for simulating the structure and mechanism of visual cortex as well as constructing the computational model of visual system. In this paper, the structures and mechanisms of the primary visual system and ventral pathway as well as the related visual computing model are fully investigated, the main works include:

Firstly, base on MCI model which is a biological vision mechanisms inspired model, we proposed sMCI and sCMCI models, which integrate with some more psychological phenomena and biological mechanisms of primary visual cortex. Therefore, they are more brain-inspired. This paper validates the models on unmanned aerial vehicle data being collected by our laboratory as well as some public datasets, and the results show that the proposed models can accelerate the speed and keep the performance of the contour detection.

Secondly, based on the hierarchical temporal memory model (HTM), this paper constructs the computational model of the ventral pathway. Based on the existing sMCI model and the connection and weight of the ventral pathway， the HTM model is improved to accord with the biological characteristics of the visual cortex. We apply it to the tasks of handwritten character recognition and illusion recognition and obtain some good results. And illusion recognition performance of HTM outperforms the result of convolutional neural network LeNet-5, which indirectly illustrates that the HTM is more conform to the mechanism of human brain.