分层计算感知模拟在目标识别技术中的应用

发布时间：2018-04-12 23:05

本文选题：目标识别 + 分层感知　；参考：《沈阳理工大学》2017年硕士论文

【摘要】：随着人类对生物视觉信息感知机制的深入了解,模拟大脑视觉感知系统以增强计算机视觉的环境感知与认知能力已经成为计算机视觉领域的研究热点之一,其主要思想是针对大脑视觉皮层中视觉信息的层次性感知过程进行模拟,以实现目标识别的目的。近年来,深度学习的发展使机器具备分层视觉特征感知能力成为可能,通过深度学习提取的特征是一种分层次的高度抽象特征,这种特征能够在目标识别等许多领域的问题求解上获得更好的效果。本文以具备稀疏连接思想和自我学习机制的卷积神经网络为框架,融入分层和仿生的思想,模拟人脑分层信息处理机制,实现显著目标的准确识别。模型首先构建多通道Gabor小波滤波器组模拟视网膜神经节细胞感受野视觉刺激响应,替代传统卷积神经网络的第一层卷积层;在此基础上,引入神经细胞间的”侧抑制”机制,采用局部响应值归一化层对神经元响应进行筛选;其中,为了模拟生物神经细胞层次间的稀疏激活性,模型引入修正非线性单元作为激活函数;并利用空间金字塔随机池化层自适应获取目标的全部有效特征,避免由于图像压缩等操作造成的特征损失;最后,模型仿照生物神经纤维膜内局部电流的随机电化学变化过程,利用随机DropConnect方法解决神经网络计算常见的过度拟合问题。所提模型通过模拟腹部通路中V1、V2、V4和IT区的功能,构建了具有分层计算感知不变性特征和具有学习、识别能力的分层计算感知模型。将该模型应用在不同规模数据集进行训练和验证,在同样识别条件下,和其他的目标识别模型对比,具有更好识别精度,识别稳定性。
[Abstract]:With the deep understanding of the mechanism of biological visual information perception, simulating the brain visual perception system to enhance the environmental perception and cognitive ability of computer vision has become one of the research hotspots in the field of computer vision.The main idea is to simulate the hierarchical perception process of visual information in the visual cortex of the brain in order to achieve the goal of target recognition.In recent years, the development of deep learning has made it possible for machines to have the perception ability of hierarchical visual features.This feature can obtain better results in many fields such as target recognition.In this paper, a convolution neural network with sparse connection and self-learning mechanism is used as a framework to simulate the hierarchical information processing mechanism of human brain and to realize the accurate recognition of significant targets by incorporating the idea of stratification and bionics.In the model, a multi-channel Gabor wavelet filter bank was constructed to simulate the visual stimulation response of retinal ganglion cells to replace the first layer of convolution layer of traditional convolutional neural networks, and on this basis, the mechanism of "lateral inhibition" between nerve cells was introduced.In order to simulate the sparse activation of biological nerve cells, a modified nonlinear element is introduced as the activation function.In order to avoid the feature loss caused by image compression and other operations, the model adaptively acquires all the effective features of the target by using the spatial pyramid random cell layer. Finally, the model mimics the stochastic electrochemical process of local currents in the membrane of biological nerve fibers.Stochastic DropConnect method is used to solve the overfitting problem in neural network computation.By simulating the functions of V1 / V2V4 and IT region in abdominal pathway, the proposed model has the characteristics of hierarchical computing perceptual invariance and the ability of learning and recognition.The model is trained and validated in different data sets. Under the same recognition conditions, compared with other target recognition models, the model has better recognition accuracy and stability.
【学位授予单位】：沈阳理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP18

【参考文献】