Dar, I., Institute of Agricultural Engineering, A.R.O., Volcani Center, Israel, Dept. of Industrial Engineering and Management, Ben-Gurion University of the Negev, Israel Edan, Y., Dept. of Industrial Engineering and Management, Ben-Gurion University of the Negev, Israel Bechar, A., Institute of Agricultural Engineering, A.R.O., Volcani Center, Israel
An adaptive path classification algorithm for a pepper greenhouse sprayer working under variable outdoor lighting conditions is described. 22 color features transformations specialized in soil-leafage discrimination extracted from the RGB and HSV 24-bit color images were created. 'Judges Vote', an innovative supervised learning methodology based on decision tree CART, was developed to classify pixels according to their color features into "Path" and "Non-Path" classes. Optimal CART feature selection was implemented by creating several single level trees. Image processing routines (including segmentation, erosion and dilution) were integrated. 12 features were selected from the original 22. Classification tests for seven random daylight videos resulted in 92% correct detection as compared to 89% correct classification obtained with regular CART classification.
An adaptive path classification algorithm for a pepper greenhouse sprayer
1
Dar, I., Institute of Agricultural Engineering, A.R.O., Volcani Center, Israel, Dept. of Industrial Engineering and Management, Ben-Gurion University of the Negev, Israel Edan, Y., Dept. of Industrial Engineering and Management, Ben-Gurion University of the Negev, Israel Bechar, A., Institute of Agricultural Engineering, A.R.O., Volcani Center, Israel
An adaptive path classification algorithm for a pepper greenhouse sprayer
An adaptive path classification algorithm for a pepper greenhouse sprayer working under variable outdoor lighting conditions is described. 22 color features transformations specialized in soil-leafage discrimination extracted from the RGB and HSV 24-bit color images were created. 'Judges Vote', an innovative supervised learning methodology based on decision tree CART, was developed to classify pixels according to their color features into "Path" and "Non-Path" classes. Optimal CART feature selection was implemented by creating several single level trees. Image processing routines (including segmentation, erosion and dilution) were integrated. 12 features were selected from the original 22. Classification tests for seven random daylight videos resulted in 92% correct detection as compared to 89% correct classification obtained with regular CART classification.