Yield Estimation

Using Feedforward Neural Networks for Color Based Grape Detection

V. Casser: “Using Feedforward Neural Networks for Color Based Grape Detection.” Technical report, University of Bonn, 2016.

Precise yield estimation is a relevant challenge for agriculture. In this paper we apply simple feedforward neural networks (FFNN) on the problem of color-based grape detection in field images, achieving an average classification rate of 93% near realtime. Our evaluation shows a detailed comparison between our FFNN approach and SVMs handled by up to date SVM implementations, revealing FFNN can slightly outperform SVMs regarding computation time while obtaining competitive results. Furthermore our results are not only competitive with state of the art results of pixelwise color-based classification in the application field of precision farming but show also new contributions by investigating the influence of using different color models and evaluating our classifiers on different lighting conditions and grape varieties. We also present a specialized software framework with extensive options for customization to manipulate and process agricultural data.

Supplementary Material

Left: images similar to those captured by field robots in different lighting scenarios. Right: software user interface.