Active Learning Based on Clustering Algorithms for the Detection of Grapes in Images

Many problems change or appear quickly over time, and it is necessary to obtain label examples for performing classification tasks with supervised learning. In agricultural classification problems where there can be variations due to seasons and other elements, it is necessary to obtain quick solutions such as for performing yield estimation. The paper details an active learning approach using clustering for minimizing the number of labeled examples in the supervised classifier, by selecting only the most relevant examples to label manually. We propose to compare k-means and spectral clustering for the creation of the trained ground truth for a multi-layer perceptron. In addition, we compare three rotation invariant texture descriptors on each color channel: 1) Brightness, Contrast & Directed Gradients Information, 2) Color Histograms, and 3) Local Binary Patterns. Finally, we assess the impact of using RGB vs. HSV as the color model in the classification performance. The evaluation is performed on two data sets of images containing grape images of different sorts. The results show that it is possible to reach an accuracy of $\approx$ 95% while decreasing the number of blocks to label from 8504 to 50. The results suggest that the proposed active learning can substantially reduce the amount of manual labor for creating the ground truth despite a limited loss in performance.