官术网_书友最值得收藏!

Finding features from an image

Let's take a look at an image of the letter X. When we feed a new image into the system, the CNN doesn't know whether the feature will match or not. Consequently, it will try to match the feature pattern everywhere across the image. This is how we build a filter:

The math logic we apply here is called convolution. To calculate the match of a feature to the portion of the image, multiply each pixel value of the feature by the value of the corresponding pixel in the image. To come to one final value, add together all of the values and divide them by the total number of pixels:

For example, if both pixels are the same color (say a value of 1) then (1) * ( 1) = (1). If not, then (-1) * (-1) = 1. In the end result, every matching pixel will have a resulting value of 1, and every mismatch will have the value of -1.

To complete this convolution process, let's move our feature grid onto the image patch. As shown in the following diagram, the 3 x 3 grid moves the 7 x 7 grid over. This forms the resulting 5 x 5 array. In the resulting grid, the values close to 1 represent strong matches, the values close to 0 represent no match, and the values close to -1 represent photographic negatives of our feature:

In the next step, we have to repeat the convolution process for all other features. This gives the filtered image—one for each of our filters. In CNN, this is known as the convolution layer, and this will be followed by additional layers that are added to it. 

This is where CNN gets into heavy computations. This example shows a simple 7 x 7 image that provides 5 x 5 as a result. However, a typical picture will be at least 128 x 128 pixels in size. The computation increases linearly with the number of features, as well as the number of pixels in each feature. 

主站蜘蛛池模板: 揭阳市| 合川市| 福鼎市| 福鼎市| 祥云县| 缙云县| 邻水| 台东市| 安义县| 益阳市| 炉霍县| 沙雅县| 沂源县| 岳阳市| 平昌县| 赤水市| 乌什县| 拉孜县| 桐梓县| 大渡口区| 确山县| 贺州市| 洪湖市| 喜德县| 电白县| 老河口市| 亳州市| 通化市| 望都县| 班玛县| 犍为县| 韶关市| 上林县| 赤壁市| 合肥市| 靖西县| 若羌县| 鹿邑县| 永仁县| 横峰县| 赣州市|