Препознавање на растителни видови преку обработка на слика и машинско учење

Abstract

Plant species recognition has recieved a great attention from the machine learning and vision communities. There are many challenges that need to be overcome in order to have a reliable plant recognition method that could be used in serious industrial applications. The problem of plant species recognition can be seen as two separate problems, based on the goal that we want to achieve. The first problem can be defined as recognizing the presence of a certain plant on an image ( retrieval problem ), and the second problem is finding a certain type of plant in an image and segmenting it from the image. In this dissertation we are reviewing state of the art achivements for both problems and give some suggestions for their improvement. The accent of the research is put on detection and segmentation of plants from images, especially from crop field images, where the main goal is detection and segmentation of unwanted weed plants. For this purpose we have recorded several datasets from seedling plantations in the Republic of Macedonia using common, commercially available cameras. We have generated two datasets, one with tobacco seedling images and the other with spinach, carrot and salad images. Both datasets were generated from images taken under variable light conditions and under slightly different heights. The tobacco seedling images were taken from the Prilep region in Macedonia, and the Salad and Carrot images were taken from the region around Negotino, Macedonia. The goal of the research was to examine the possibility of using different machine vision and learning algorithms for the purpose of plant weed segmentation from the images. This approach would be the first step of the design and implementation of a low cost, machine learning based sensor system, that could be used for weed control automation. The results of this dissertation show that there is significant improvement in the weed detection by using the novel machine learning and deep learning methods, however there are still some challenges that need to be overcome so that common RGB cameras can be used for weed detection. We also present a novel user interface for data generation and annotation that can be used by farmers for specialized model training for specific plant types. This approach could be used for plant-weed segmentation datasets, but it could also be used for dataset generation in other types of segmentation problems.