Improving the segmentation for weed recognition applications based on standard RGB cameras using optical filters

Within precision agriculture we have seen an increase in the utilization of computer vision systems both in academia and in commercial products. Within the agricultural industry computer vision is primarily used for tractor and machine guidance whereas in academia it is commonly used for detecting diseases, weeds and fungus. A common method for interpretation is based on the leaf shape. However in order to reliably achieve a good description of the shape a good segmentation is required.
The excess green index is one of the most common methods for green vegetation segmentation within agriculture. This method utilizes that most vegetation reflects more green light than blue and red. As silicon based image sensors is also sensitive to near-infrared light a typical rgb-camera will have a filter in place to block the near-infrared light. When using excess green the ideal filter would be a sinc-filter following a rectangular function. However the filter in place is selected for best mimicking the spectral sensitivity of the human vision, the cut-off is therefore neither sharp nor blocks completely.
In this work we show that by replacing the IR filter with a more carefully selected IR filter matched for green vegetation segmentation we are able to attain a significantly improved segmentation under controlled illumination.