Computer Vision
Give a robot to "see" the function is one of the research question "machine vision" of the discipline. This field is very broad, including not only general-purpose technology, but also a large number of proprietary technology - such as NLP, fingerprint, photo interpretation and robot control and so on. Here are some only computer vision concepts.
introduction
First, computer vision is set on the original photosensitive, generating an image of a scene. This is a camera image of a scene in a field of view by perspective projection lens, and a photoelectric conversion element which is a two-dimensional, time-varying luminance matrix image I (x, y, t), where x, y is the position of the photovoltaic element in the array, t is time (for colored visual matrix is required to represent such three primary colors). A visual response from the agent to be guided to generate a model of the scene icon, or by processing a set of features of this matrix, so that he can direct a computing operation.
We want to get the kind of information depends on the agent's purpose and mission. To make a agent safely through a chaotic environment, this agent must understand the position in which the properties of objects, boundaries, and its path through the path of the surface. Perhaps the agent should also have to predict the future changes can all change from time to time in accordance with the above information. Or obtained from a plurality of images and such information difficult, therefore, can only be given an overview of such techniques.
Steering a car
SR agent In some applications, the neural network can be used to direct the image into luminance matrix operation. One prominent example is used to driving a car ALVINN system.
The first hidden layer network unit 5, the second layer 30 output units, all of the above units are sigmoid unit. The output unit is controlled by the height of the car linearly arranged. If this output unit output unit near a top of the queue is higher than most other output unit, the vehicle travels to the left; if this output unit near the bottom end of an output queue unit higher than most other output unit, the vehicle travels rightward .
This system had improved "in the air (on-the-fly)" training methods to spread by the real driver's car, the actual driving angles are correct sign as the corresponding input. Back-propagation network by way of incremental training, so that he can drive the angle of the driver designated to respond to each visual pattern that occurs when the actual driving of the vehicle.
Two-stage machine vision
- Image processing stage to convert the original image into a scene more suitable for image analysis.
- 图像处理包括降噪、增强边缘和寻找图像区域等不同的滤波操作。
- 景物分析主要试图从已处理的图像中产生一个对原始场景的图标描述或基于特征的描述，并提供agent所处环境中与特定任务有关的信息。
图像处理
1. 平均法
假设初始图像可表达为一个m*n的数组I(x,y),我们称之为“图像亮度数组”。他把图像平面分成许多被称为“像素（pixel）”的单元。这些数字表示这幅图相中某点的光亮度，图像中一些不规则之处可通过求平均数的方法得以平滑。这种滑动并求和的操作称为“卷积”。如果我们的得到的数组十二进制（1或0），那么就必须把这些加权总和和一个阈值比较。平均法不仅将压缩孤立的噪音点，而且将减小图像的卷曲度（crispness），放弃那些微不足道的图像元素。
有时，我们把加权函数W(x,y)的值在x和y构成的长方形内看做1，长方形之外看做0.长方形大小决定平滑度，长方形越大平滑度越高。下图展示了一个求平均数操作是如何让对一个二进制图像先用一个长方形平滑函数平滑，然后将其与阈值比较来进行操作的。
我们发现这个平滑操作加粗了宽线，去除了窄线和微小细节。
2. 边缘增强
如前所述，计算机视觉常常设计图像边缘的提取，然后用这些边缘来把图像转换成某种线条图形。获取轮廓的方法之一是先增强图像中的边界和边缘，边缘可以是图像个部分之间的任意边界。
我们可以通过在以为图像上卷积一个位于垂直线上的、一半为负一半为正的窗口来增强这些图像的边缘强度。
3. 边缘增强和平均法的结合
还有其他的变化比拉普拉斯变换更好，其中突出的有：Canny变换、Sobel变化、Hueckel变换等。
4. 区域查找
首先，我们必须定义什么是图像的一个区域。一个区域就是满足一下特性的相互连接的像素：
下图运用了亮度差别不超过1个单元这个同质的特性。当无需再进行分割时，可以合并那些满足此同质特性的相邻的候选区域。
场景分析
在用以上技术对图像进行处理后，我们力图从中获取所需有关场景的信息。计算机视觉的这个阶段被称为“场景分析”。
1. The explanation of the image lines and curves
When the scene contains a known linear objects to be analyzed, where the key step is assumed that the image lines. It may be generated by using the straight line segments and a region of the boundary edge or fitting techniques. The figure is the interpretation of a room.
2. Based on Vision Model
