Shulou(Shulou.com)06/03 Report--

Editor to share with you what the convolution kernel in CNN is, I believe most people do not know much about it, so share this article for your reference, I hope you can learn a lot after reading this article, let's go to know it!

Convolution kernel is that in image processing, given an input image, the pixels in a small area of the input image are weighted and averaged to become each corresponding pixel in the output image, in which the weight is defined by a function, which is called convolution kernel.

The significance of convolution kernel

Convolution kernel is not a new thing in image processing. A series of filter operators, such as Sobel operator, have been used in edge detection all the time, but they used to be called Filter. Students who do image processing should have an impression.

One of the properties of convolution kernels is locality. That is, it only pays attention to the local features, and the degree of the local depends on the size of the convolution kernel. For example, the essence of edge detection using Sobel operator is to compare the similarity of the adjacent pixels of the image.

The meaning of convolution kernel can also be understood from another point of view. Students who have studied signal processing should remember that time domain convolution corresponds to frequency domain multiplication. Therefore, the convolution of the original image and the convolution kernel is actually the selection of frequency domain information. For example, the edges and contours in the image belong to high-frequency information, and the comprehensive consideration of the intensity of a certain region in the image belongs to low-frequency information. In traditional image processing, this is an important aspect to guide the design of convolution kernel.

Convolution Kernel in CNN

The convolution kernel in CNN is no different from the traditional convolution kernel in nature. Still take the image as an example, the convolution kernel is convoluted with the image blocks at different positions of the input in turn to get the output.

At the same time, CNN has some unique features, such as various definitions:

CNN can be seen as a simplified form of DNN, that is, each weight in the convolution kernel can be regarded as DNN, and like DNN, there will be one more parameter offset.

When a convolution kernel is convoluted in different regions from Input, its parameters are fixed. In the framework of DNN, it is understood that for neurons in the same layer of Layer, their sum is the same, but the connected nodes are changing. So in CNN, this is called shared weight offset.

In CNN, convolution kernels may be high-dimensional. If the input is dimensional, then the general convolution kernel will choose to be dimensional, that is, consistent with the input Depth.

Most importantly, in CNN, the weight of the convolution core does not need to be designed in advance, but is optimized using GD like DNN, we just need to initialize.

As mentioned above, in fact, the convolution kernel convolution will get some features of the original image (such as edge information), so in CNN, the Layer obtained by convolution kernel convolution is called feature graph.

In a word, there are multiple convolution kernels between the two layers of CNN in order to learn the different features of Input and get multiple feature graphs. And because the parameters in the convolution kernel are optimized by GD rather than set by us, initialization is particularly important.

The above is all the content of the article "what is the convolution kernel in CNN". Thank you for reading! I believe we all have a certain understanding, hope to share the content to help you, if you want to learn more knowledge, welcome to follow the industry information channel!

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.