WAVELET-BASED EFFECTIVE IMAGES DOWNSCALING IN NEURAL NETWORKS

Based on discrete wavelet transform, several blocks for images downscaling in computer vision models were implemented. The blocks were tested with ResNetV2-50 and MobileNetV2 models on Flowers dataset. With small increase in number of models’ parameters and close results in terms of metric the changes allowed to reduce number of training epochs by 34 % and VRAM requirements by 18 %. Due to the implementation details the blocks suggested can be used as a replacement of layers responsible for images downscaling in models for other tasks to save computation resources and speed up training process. In the blocks developed standard operations of addition and multiplication are used for evaluation of wavelet transform, which allows a simple export of trained models into other formats.