Contents lists available at ScienceDirect Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag Original papers A pattern recognition strategy for visual grape bunch detection in vineyards Rodrigo Pérez-Zavala a , Miguel Torres-Torriti a, ⁎ , Fernando Auat Cheein b , Giancarlo Troni c a Dept. of Electrical Engineering, Pontificia Universidad Católica de Chile, Chile b Dept. of Electronic Engineering, Universidad Técnica Federico Santa María, Chile c Dept. of Mechanical Engineering, Pontificia Universidad Católica de Chile, Chile ARTICLE INFO Keywords: Grape bunch detection Grape recognition Precision viticulture Histogram of oriented gradients Local binary pattern Support vector machine ABSTRACT Automating grapevine growth monitoring, spraying, leaf thinning and harvesting tasks, as well as improving yield estimation and plant phenotyping, requires reliable methods for detecting grape bunches across different vineyard environmental and plant variety conditions, in which illumination, occlusions, colors and contrast are the main challenges to computer vision techniques. This work presents a method that employs visible spectrum cameras for robust grape berries recognition and grape bunch detection that does not require artificial illumi- nation nor is limited to red or purple grape varieties. The proposed approach relies on shape and texture in- formation together with a strategy to separate regions of clustered pixels into grape bunches. The approach employs histograms of oriented gradients (HOG) as shape descriptor and local binary patterns (LBP) to obtain texture information. A review of the existing methods and comparative analysis of different feature vectors (DAISY, DSIFT, HOG, LBP) and support vector classifiers (SVM-RBF, SVDD) is also presented. Datasets from four countries containing 163 images of different grapevine varieties acquired under different vineyard illumination and occlusion levels were employed to assess the approach. Grapes bunches are detected with an average precision of 88.61% and average recall of 80.34%. Single berries are detected with precision rates above 99% and recall rates between 84.0% and 92.5% on average. The proposed approach should facilitate the estimation of yield, crop thinning measurements and the computation of leaf removal indicators, as well as the im- plementation guidance strategies for precise robotic harvesters. 1. Introduction The industry of fresh table and wine grapes represent a major eco- nomic activity worldwide, being the fruit crop with highest value, with a market size of approximately 70 billion dollars; see Fig. 1,(FAO-OIV, 2016). Chile is worlds leading fresh table grape exporter (USDA, 2014), and one of the major wine producers, with 5% of the global production (OIV, 2017). Statistics show slightly increasing trends in surface and production during recent years. However, the viticulture industry is facing difficulties due to shortage of qualified fieldworkers and in- creasing labor costs (Canadian Agricultural Human Resource Council, 2016; Quackenbush, 2017), which affect productivity, quality, and timely harvesting. Additionally, agricultural tasks manually done are often time consuming, inaccurate and subjectively influenced or biased by workers (Nuske et al., 2014). These challenges motivate the devel- opment of new technologies that rely on novel sensors and robotics to ensure productivity, quality and economic competitiveness. Robotic systems in agricultural applications involve three main components: mobility algorithms and hardware (guidance and mapping), perception subsystems, and end effector actuation mechan- isms. This work focuses on the perception stage for detection and re- cognition of grape bunches in the field. For any robotic or automation process, the sensing stage is crucial for the correct performance of the unit. The detection and localization of grape bunches in vineyards has been an important challenge that still has not been completely solved. The automatic recognition of grapes and grape bunches using ma- chine vision could be employed to automate, manage and optimize current agricultural tasks, such as harvesting (Luo et al., 2016), spraying (Berenstein et al., 2010), grape counting for yield estimation models (Diago et al., 2015; Dunn and Martin, 2004; Liu and Whitty, 2015; Nuske et al., 2014), evaluating grape quality, size and grapevine phenotyping (Roscher et al., 2014; Yandun-Narvaez et al., 2017), de- tecting disease in clusters, predicting harvest time, quantifying and standardizing crop thinning and basal leaf removal tasks. While several of the grape detectors developed recently have achieved good performance scores, improvements in grape bunch de- tection are still possible. The main contribution of this work is the de- velopment of a grape bunch detection approach, and not just a grape https://doi.org/10.1016/j.compag.2018.05.019 Received 24 September 2017; Received in revised form 12 May 2018; Accepted 14 May 2018 ⁎ Corresponding author. E-mail address: mtorrest@ing.puc.cl (M. Torres-Torriti). Computers and Electronics in Agriculture 151 (2018) 136–149 0168-1699/ © 2018 Elsevier B.V. All rights reserved. T