Style Transfer as a Way to Improve the Generalization Ability of a Neural Network in an Object Detection Task
Denis Konstantinovich Karachev, Sergey Evgenievich Shtekhin, Vladimir Sergeevich Tarasyan, Ilya Urevich Smolin, Maksim Vladimirovich Isakov · Proceedings of the Institute for System Programming of RAS · 2023
This paper proposes the implementation of a neural network training approach for object detection, using augmentation - style transfer. This method improves the generalization ability of the neural network to determine the location of objects in the image by improving the interaction with low-level features such as textures, different colors and small changes in shapes. The effectiveness of the method is experimentally proved and the numerical values of the object detection metrics are demonstrated on several datasets with different classes. The application of augmentation is proposed using an unused before neural network architecture capable of carrying an arbitrary number of styles. The peculiarity of the approach is also that the weights of the neural network for styling are frozen and it is added to the graph of the detection network, which allows augmentation speed.