Pattern Recognition on Radar Images Using Augmentation
Nikita Andreevich Andriyanov, Danila Andriyanov · 2020 Ural Symposium on Biomedical Engineering, Radioelectronics and Information Technology (USBEREIT) · 2020
The text describes different models of convolutional neural networks (CNN) using to multiclass recognition of objects on radar images. It was suggested to use data augmentation methods because of limited size of source database. The main augmentation types are as follows: additive white Gaussian noise, blur, motion, shift, scale, grid distortion, brightness changing, flip, rotate and etc. A comparative analysis of recognition accuracy for the CNN based on simplest gradient descent (SGD) and adaptive moment estimation (ADAM) optimization techniques with different training duration time (number of epochs) and the neural network architecture is performed. After applied standard transformations to source images the size of source samples was increased. One of CNN provided excellent accuracy on the all images in dataset despite of its belonging to the training, validation or test sample. The developed CNN allows recognizing planes, roads and ships with quite good accuracy when the task condition is to recognize only one main pattern on the image.