Improving the Quality of X-ray Images of Seeds in Smart Farming Using Deep Learning

This work is devoted to the problem of food security and seed quality assessment in the task of automation and optimization of technological decision-making processes in agriculture. In particular, the application of neural network methods to improve the image resolution of X-ray images of seeds to increase the accuracy of their subsequent analysis is considered. The authors proposed a procedure for collecting, processing and augmenting a training set of X-ray images of grain crop seeds. Five neural network models of super-resolution, such as SRCNN, EDSR, SRGAN, ESRGAN and SwinIR, were implemented, trained, and adapted to the specifics of the subject area (quality seed), and experiments were conducted to fine-tune them. The experiment showed that the ESRGAN model has the best values of objective metrics (PSNR = 28.83 and SSIM = 0.80). The paper also shows examples of generated images, which can later be used to solve the problem of detecting the seed quality and classifying types of defects from X-ray images. The resulting solution, in addition to improving the resolution of X-ray analysis of grain crop seeds, will reduce the sensitivity of operators of specialized stations to X-ray radiation when creating and processing images.