Image Classification Based on a Weighted CNN Model
Hao Peng, Jian Zhou, Shenglan Liu · 2022 World Automation Congress (WAC) · 2022
This paper explores a deeper network architecture diagram, which can be viewed as an enhanced version of AlexNet. In the design of the winder, a smaller winder core is used, which verifies that the small size of the winder core in the deep network can not only reduce the parameters, but also achieve better results. The concept of Inception was introduced and the idea of "widening" was designed under the prevailing model of "deepening" at that time. Moreover, two weighted Softmax branches can be conducted, which consist of two functions. One is to avoid the disappearance of the gradient, and is used for conducting the gradient forward. If one of the layers is 0 in the back propagation, the chain is 0. The second is to utilize the output layer of the middle layer as a classification to analyze the significant role in the model fusion. They decompose the convolution of the filter size NXN into a combination of 1xn and nx1 convolution. For example, 3 by 3 convolution is equivalent to first performing 1 by 3 convolution, and then performing 3 by 1 convolution on its output. They found that this method is 33% cheaper than a single 3x3 convolution.