This paper considers the influence of different color space resampling methods on the training quality of neural networks of fully connected and convolutional architectures. It also studies the use of the Haar and Daubechies D4 wavelet transforms and wavelet filtering before space resampling in order to improve the quality of neural networks after color space resampling. The resampling methods considered are a bitwise shift by 3 bits used to switch from the RGB format to the R5G5B5 format and a strong compression method with splitting the color space into two, three and four intervals, allowing storing the pixel color value in two bits. A dataset of black-and-white images of handwritten digits is used to test the influence of the resampling algorithms on the quality of neural networks. At the end of this paper, the accuracy metric values are given for two architectures using two types of wavelet filtering and eighteen types of resampling. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.