The yield of agricultural crops is directly related to the quality of the seed. One of the ways to improve the quality is to sort seeds by physical and optical properties (color, shade of color, shape, size and structure) using grain cleaning units (photoseparators or color sorters). Currently, two main approaches are used to create sorting algorithms: generalized statistical and neural networks. The benefits of a neural-based sorting approach include the absence of any restrictions on training series; there is no need to pre-examine the nature of the data and adjust sorting algorithm. Nowadays, neural networks are better than statistical methods, especially in the case of deep neural networks. The disadvantages of neural networks are in the high computational complexity of obtaining results and huge size of the training series, sufficient to achieve high accuracy in making reliable decisions. This work discusses the creation of a convolutional neural network architecture that allows dividing the input flow of wheat seeds into two classes: "good" and "bad" (with flaws in shape and color). The implementation of convolutional neural network on FPGA using the Xilinx System Generator for DSP and Matlab / Simulink package is also considered. The results of a convolutional neural network training, assessment of the accuracy of seed correct classification, and verification of the convolutional neural network hardware FPGAs implementation are presented.
|