Detection of grapes in natural environment using support vector machine classifier

Abstract

The detection of grapes in real scene images is a serious task solved by researches dealing with precision viticulture. The detection of wine grapes of red varieties is a well mastered problem; however, the detection of white varieties still poses challenges. In this paper, four detectors for white wine grapes detection are introduced and evaluated. The detectors are based on support vector machines and they differ in kernels and features used for image representation. Namely, the pixel intensities and histogram of oriented gradients (HOG) are used for the representation of images. Radial basis functions and linear kernels are applied. The detectors based on the HOG feature have proven to be very efficient. Their average recognition accuracy by cross-validation was 98.23% and 98.96%, respectively. Furthermore, they show very good performance for other cross-validation metrics. Their average precision is 0.978 and 0.985, respectively; their average recall is 0.987 and 0.994, respectively. The detectors were also verified on test sets with positive samples affected by rotation distortion, and moreover on image sections of a real scene photo, in both cases with good results. Moreover, the detectors do not require any artificial lighting and they can work under different light conditions.