3.2.2. ResNet residual theory

In order to solve the gradient disappearance problem, He et al. (2016) proposed ResNet. Residual Network is proven to handle the vanishing gradient and effective feature learning better. This study uses the residual neural network (ResNet-50) as the base framework. ResNet-50 has 50 layers of CNNs, as well as a MaxPool and a fully connected layer with a softmax layer. resNet builds the network by stacking the remaining connections on top of each other. Even when the architecture becomes more complex, the ResNets model remains as efficient as before, making it a better choice than other architecture models(Praveen et al., 2022). The most important idea of ResNet is that the X output from the previous layer, after the convolution calculation of this layer to get the post F(X) , the X and F(X) will be added to get H(X) . The purpose of this is that when even if the. gradient tends to 0, the item X will still leave 1, cleverly avoiding the gradient during the backpropagation. The residual structure of the core in ResNet is shown in Figure 6 .

ResNet block.

Although the ResNet network is good for avoiding overfitting, there is still room for improvement. At the beginning of the design of the ResNet model, the model mainly consists of four Stages containing different numbers of Bottlenecks, and the ratio of the number of Bottlenecks in the Stages is largely proposed empirically, for example, the ratio of ResNet50 is 3:4:6:3, the ratio of ResNet101 is 3:4:23:3, and the ratio of ResNet152 is 3:8:36:3. It can be seen that there are also more excellent computational ratios of the number of Bottleneck in the Stage, which makes the model performance more excellent.