Sistema de visão computacional para reconhecimento e classificação de padrões de famílias de plantas invasoras

Abstract

Computer Vision, in addition to involving pattern recognition and object classification techniques, has been characterized as an emerging field of fundamental importance in the context of intelligent computing. Its application has involved different areas of knowledge, which includes the agricultural area. Since its beginnings, which involved the interpretation of manuscripts and typed texts, it currently plays a fundamental role in the production of food and sustainable energy. This research seeks to use and develop Computer Vision for the identification and localized recognition of invasive plants, which require control in order to properly develop agricultural crops. With this in perspective, algorithms and concepts of Computer Vision and Internet of Things can be used to help with this task. Then, the construction of a stereo system capable of acquiring digital images in the field, obtaining depth information, segmenting, extracting features and classifying plants through supervised machine learning methods is carried out. Thus, the recognition of patterns of invasive plant families can help in the rational use of inputs and minimize environmental impacts, entering the list of precision agriculture technologies. For the development of the system, embedded control hardware and software were validated, allowing real-time action and wireless communication. Signal and image processing actions were applied to two databases referring to crops that are impacted by the presence of invasive plants, maize (Zea mays) and groundnut (Arachis hypogaea). For the pre-processing step, camera calibration parameters were obtained for stereo image rectification. For the segmentation step, thresholding in the HSV color space and morphological operations were performed. Accuracy and overlapping area metrics were used to validate the process. The disparity maps were obtained from the evaluation of local and semi-global matching algorithms. For the feature extraction step, descriptors of Local Binary Patterns and Haralick moments were applied. For the classification of invasive plants, classifiers based on Support Vector Machines were used. The developed system proved to be capable of generating depth information and plant classification, which contributes to decision-making and the application of different input rates. The results obtained prove the effectiveness of the method developed for recognition, qualification and classification for decision support in the control of families of invasive plants.