Resizing images to improve clustering

Markeev Maksim

The purpose of this article is to investigate the dependence of the accuracy of neural networks (image clustering) on the size and proportions of images at the input of the neural network itself. Modern neural networks are used for image recognition, and they do it with great accuracy, sometimes even more accurately than humans. The problem is that the images themselves are not perfect. To improve the quality of recognition, the same image is recognized in different scales, rotations and mirroring. The work of this technique was tested on the partitioning of images into 2 clusters "cats" and "dogs". Studies have shown that the best results are obtained by zooming in the image by 30% at a height of 286 - 346 pixels and a width of 272 - 383 pixels for the convolution neural network, which was trained on a size of 224 x 224. The results may be different on different data sets, so calibration is required in each case.

Keywords:neural networks, clustering, image resizing, artificial intelligence, Keras, TensorFlow