Feature Selection: Evaluation, Application, and Small Sample Performance
IEEE Transactions on Pattern Analysis and Machine Intelligence
Selection of relevant features and examples in machine learning
Artificial Intelligence - Special issue on relevance
Learning Decision Trees for Unbalanced Data
ECML PKDD '08 Proceedings of the 2008 European Conference on Machine Learning and Knowledge Discovery in Databases - Part I
An improvement on floating search algorithms for feature subset selection
Pattern Recognition
Intelligent Service Robotics
Extraction of Litchi Stem Based on Computer Vision under Natural Scene
CDCIEM '11 Proceedings of the 2011 International Conference on Computer Distributed Control and Intelligent Environmental Monitoring
Determination of the number of green apples in RGB images recorded in orchards
Computers and Electronics in Agriculture
Computer vision for fruit harvesting robots state of the art and challenges ahead
International Journal of Computational Vision and Robotics
Classification of plant structures from uncalibrated image sequences
WACV '12 Proceedings of the 2012 IEEE Workshop on the Applications of Computer Vision
| Hi-index | 0.00 |
Sweet-pepper plant parts should be distinguished to construct an obstacle map to plan collision-free motion for a harvesting manipulator. Objectives were to segment vegetation from the background; to segment non-vegetation objects; to construct a classifier robust to variation among scenes; and to classify vegetation primarily into soft (top of a leaf, bottom of leaf and petiole) and hard obstacles (stem and fruit) and secondarily into five plant parts: stem, top of a leaf, bottom of a leaf, fruit and petiole. A multi-spectral system with artificial lighting was developed to mitigate disturbances caused by natural lighting conditions. The background was successfully segmented from vegetation using a threshold in a near-infrared wavelength (900nm). Non-vegetation objects occurring in the scene, including drippers, pots, sticks, construction elements and support wires, were removed using a threshold in the blue wavelength (447nm). Vegetation was classified, using a Classification and Regression Trees (CART) classifier trained with 46 pixel-based features. The Normalized Difference Index features were the strongest as selected by a Sequential Floating Forward Selection algorithm. A new robust-and-balanced accuracy performance measure P"R"o"b was introduced for CART pruning and feature selection. Use of P"R"o"b rendered the classifier more robust to variation among scenes because standard deviation among scenes reduced 59% for hard obstacles and 43% for soft obstacles compared with balanced accuracy. Two approaches were derived to classify vegetation: Approach A was based on hard vs. soft obstacle classification and Approach B was based on separability of classes. Approach A (P"R"o"b=58.9) performed slightly better than Approach B (P"R"o"b=56.1). For Approach A, mean true-positive detection rate (standard deviation) among scenes was 59.2 (7.1)% for hard obstacles, 91.5 (4.0)% for soft obstacles, 40.0 (12.4)% for stems, 78.7 (16.0)% for top of a leaf, 68.5 (11.4)% for bottom of a leaf, 54.5 (9.9)% for fruit and 49.5 (13.6)% for petiole. These results are insufficient to construct an accurate obstacle map and suggestions for improvements are described. Nevertheless, this is the first study that reports quantitative performance for classification of several plant parts under varying lighting conditions.