Visual identification method based on deep convolutional neural network model-regeneration network

Abstract

According to a visual identification method based on a deep convolutional neural network model-regeneration network, a novel deep convolutional neural network model is built by adopting a regenerationmodule of a regeneration mechanism, the introduced regeneration mechanism carries out regeneration and reconstruction on neurons cut off and dead by a ReLU function, and the realization process of the regeneration mechanism comprises the following steps: in the regeneration network, as a regeneration module. Firstly, an input x of a regeneration module is a feature map obtained by an upper convolution layer, and the x is input into a traditional ReLU function to obtain an activated feature map x1, so that neurons with positive values are screened out, and neurons with negative values are cutoff; meanwhile, the input x is negated, and the xis input into a ReLU function in parallel to obtain activated feature mapping x2 *, so that neurons with negative values are screened out, and neuronswith positive values are cut off; and the neurons with negative values are screened, inverse convolution operation is performed on the neurons, and channel cascading with positive values is performedto obtain a regeneration process of the negative neurons.

Description

technical field [0001] The present invention relates to a novel and high-performance deep convolutional neural network model, which belongs to artificial intelligence, especially the deep convolutional neural network model-reborn network (RebornNet) visual recognition method, the field of deep learning and computer vision, and can be used for images Tasks such as recognition and image classification can also be used as benchmark models for tasks such as object detection, semantic segmentation, image generation, and style transfer. Background technique [0002] Convolutional Neural Network (CNN) is a feed-forward neural network, which has excellent performance in large-scale image processing. The essence of the convolution operation: the input image (input volume) is composed of many slices in the depth direction. For one of the slices, it can correspond to many neurons. The weight of the neurons is in the form of a convolution kernel, that is, a square filter. (filter) (suc...

AI Technical Summary

Problems solved by technology

However, mapping the negative input to 0 makes the ReLU function have a "death characteristic", so that the gradient of the neuron node with an output of 0 is 0, where the weight will not be updated during the backpropagation process, resulting in the death of the neuron

[0013] Parameter initialization or too large a learning rate may cause this problem

At the same time, negative input may also contain useful information, and the ReLU function also causes the input information to not be fully and effectively utilized.

In addition, studies have shown that the shallow layer of the deep neural network is equally sensitive to positive and negative phase inputs, and the truncation of the negative value by the ReLU function will lead to a decrease in the utilization of the convolution kernel channel, which is called "network parameter compensation". Phenomenon"

At the same time, the basic traditional convolutional neural network has symmetry, which is not conducive to the further improvement of the network representation ability.

Images