An algorithm was presented to detect size of berries on grapes bunch. Firstly, the contour curves of the grapes bunch were extracted by contour tracking. Then concave points were detected by rotating the curve continuously and searching local extremum points so that the curves were divided into circular arcs, in which each arc corresponds to one berry of the grapes. Finally, the least-square curve fitting method was applied to calculate the radius of the berries. Experiments of Jufeng grapes showed that by using the presented method, about 35% berries on a grapes bunch could be recognized. In the estimation of average berry radius of a bunch, the average error was 0.61mm, the maximum error was 1.69mm and berry size grading accuracy was 72.7%.