Modulation signal recognition based on lightweight complex residual attention neural network

Research on signal modulation identification method based on residual neural network

“…In addition, considering that the signal feature distribution is different from the image pixel values, and the feature values are both positive and negative, the SeLU function is used after the normalization layer instead of the traditional activation function ReLU to increase the nonlinearity of the model, while retaining the negative features of the signal as much as possible [1]. Each Residual Stack unit includes a convolution kernel of size (1, 1), which is used to do the computation in channel dimension, in the first Residual Stack unit, due to the input data format of (2, 128) for the two IQ data, set the size of the convolution kernel inside the residual module of the Residual Unit are (2,5). The number of convolution kernels is 32, and after a maxpool of size (2, 2) the dimension becomes (1,64), at this time the dimension of the convolution kernels inside the Residual Unit residual module in the next Residual Stack unit should not be larger than the input data, so the size of the convolution kernels inside the second Residual Stack unit are both (1,5), and the number of convolution kernels is 32, which facilitates the input data format of (2, 128) two-way IQ data.…”

“…Each Residual Stack unit includes a convolution kernel of size (1, 1), which is used to do the computation in channel dimension, in the first Residual Stack unit, due to the input data format of (2, 128) for the two IQ data, set the size of the convolution kernel inside the residual module of the Residual Unit are (2,5). The number of convolution kernels is 32, and after a maxpool of size (2, 2) the dimension becomes (1,64), at this time the dimension of the convolution kernels inside the Residual Unit residual module in the next Residual Stack unit should not be larger than the input data, so the size of the convolution kernels inside the second Residual Stack unit are both (1,5), and the number of convolution kernels is 32, which facilitates the input data format of (2, 128) two-way IQ data. The number of convolutional kernels is 32, which is favorable to reduce the number of neural network parameters.…”

“…In order to solve the defects such as gradient dispersion of convolutional neural networks, in 2018, the authors in the literature [4] proposed a modulation recognition method for radio signals based on Deep Residual Neural Network, which introduces multiple Residual Blocks (Residual) structures based on the use of convolutional neural networks in the literature [3] to improve the model. Blocks) structure to improve the model recognition accuracy and generalization ability, the experimental results show that the number of parameters of the modulation recognition model based on residual neural network is 2.3×105, and the recognition accuracy is up to 94% when the training dataset is 740,000.In 2020, S. Hu et al [5] proposed a classifier based on deep neural networks (DNNs) to classify the signal modulation recognition. In the above literature study, it is found that even a simple CNN neural network modulation recognition model has a parameter count of up to 2.8×106 and requires 40 minutes of training, which is unaffordable in a communication environment where the actual resource conditions are limited and require constant training and adaptation to different conditions.…”

Research on similar handwritten character feature extraction algorithm by introducing deformable convolution

“…In addition, the application of attention mechanisms and deformable convolution(DC) has received attention in improving image local feature extraction. In the literature [17][18], the attention mechanism is fused into the convolutional structure to enhance the model's feature extraction of target local region locations, thus further improving recognition and classification accuracy. DAI [19] proposed deformable convolution, which realizes the adaptive change of convolution kernel shape and improves image feature extraction.…”