Program for automatic clustering of numerous (over 1000) images of pollen grains obtained using an optical microscope in Python, for learning the LeNet convolutional neural network

Abstract

A method of images classification of pollen grains at the preprocessing sampling stage for learning of the LeNet convolutional neural network is considered. The Python program was developed for cropping an image of pollen obtained using an optical microscope, ImageSplitter by which an image of polyfluoric honey was framed at 1131 images 128 pixels * 128 pixels in size with a sliding window step of 32 pixels. The primary Python classification program was developed which allows sorting out empty images from non-empty ones through the task conditions and intervals of the degree of fullness in black. A study on the possibility of sorting non-empty images into three classes according to the percentage of fullness in black was conducted. The proportion of black points among all Perimeter total (i.e., the criterion of the area occupied by a grain of a frame) is calculated; the criterion of the frame's perimeter grain is Perimeter black and in percentage the Perimeter percentage. The inverted threshold binarization was performed to select objects. The following results were obtained: in the interval from 0 to 30, the program did not display any images, i.e., the folder output_result_0-30 was empty; in the interval from 31-60, the program has classified 4 images, i.e., there were 4 images in the folder output_result_31-60; between 61-99, the program has classified 130 images.