Image-based recognition framework for robotic weed control systems
Tsampikos Kounalakis,Georgios A. Triantafyllidis,Lazaros Nalpantidis
DOI: https://doi.org/10.1007/s11042-017-5337-y
IF: 2.577
2017-11-09
Multimedia Tools and Applications
Abstract:In this paper, we introduce a novel and efficient image-based weed recognition system for the weed control problem of Broad-leaved Dock (Rumex obtusifolius L.). Our proposed weed recognition system is developed using a framework, that allows the examination of the affects for various image resolutions in detection and recognition accuracy. Moreover, it includes state-of-the-art object/image categorization processes such as feature detection and extraction, codebook learning, feature encoding, image representation and classification. The efficiency of those processes have been improved and optimized by introducing methodologies, techniques and system parameters specially tailored for the goal of weed recognition. Through an exhaustive optimization process, which is presented as our experimental evaluation, we conclude to a weed recognition system that uses an image input resolution of 200 ×150, SURF features over dense feature extraction, an optimized Gaussian Mixture Model based codebook combined with Fisher encoding, using a two level image representation. The resulting image representation vectors are classified using a linear classifier. This system is experimentally shown to yield state-of-the-art recognition accuracy of 89.09% in the examined dataset. Our proposed system is also experimentally shown to comply with the specifications of the examined applications since it provides low false-positive results of 4.38%. As a result, the proposed framework can be efficiently used in weed control robots for precision farming applications.
computer science, information systems, theory & methods,engineering, electrical & electronic, software engineering