Development of a Computer Vision System for an Optical Sorting Robot

The use of sorting robots at some stages of agricultural production seems quite promising. In addition to the high-tech design, the most important element of such robots may be a pattern recognition system, which in turn, in addition to optical devices, includes an intelligent decision support system. This paper presents the development of a computer vision system for recognizing suitable samples of grain crops for an optical sorting robot. The aim of the work is to create a reliable and efficient algorithm that can accurately determine the presence and characteristics of damage in images of wheat, oats and peas. Modern machine learning methods, such as convolutional neural networks, were implemented to train the recognition model. The development process included collecting and preparing training data, selecting and setting up the neural network architecture, as well as testing and optimizing the algorithm. A comparison of computer vision libraries YOLO, FASTER R-CNN, VISSL, OpenCV was carried out. The resulting system demonstrated high accuracy in recognizing defects and morphological features of seeds in test images, with an accuracy of up to 87%. The developed system can be used in optical sorting robots and various mechatronic applications related to automation of agricultural processes, product quality analysis, robotic phenotyping devices, as well as in seed quality control systems and in intelligent control systems for agricultural production processes in crop production.