Training of deep neural networks for classification of biological objects based on spectral measurements

Abstract

The possibility of using deep neural networks for processing multispectral measurements in the classification of biological objects with specific pathologies is estimated. Convolutional neural networks accepting as an input multiple spectral functions (amplitude and frequency dependencies) retrieved simultaneously for each object using various spectral methods (reflection, transmission) or different wavelength ranges. The structure of a deep network with two convolutional layers and two fully connected layers for classifying wheat seeds in order to identify grains affected by fungal diseases and to identify the variety of the seeds is described. A method for improving the quality of network learning based on applying transfer learning method after training on artificially generated data is proposed. A comparison of the classification results obtained for reflection and transmission spectra of elements of grain mixtures with traditional deep learning approaches (for each spectral method separately), with learning on data for two spectral methods (simultaneous processing of two data channels representing the transmission and reflection spectra), and with the proposed transfer learning-based method is provided.