Leaf recognition using multi-layer perceptron

Abstract

In this work, we employ multi-layer feed forward network algorithm for leaf recognition using its shape, color and vein features. The leaf images are pre-processed and segmented. The shape features like area, convex area, diameter, length, width, perimeter, eccentricity, solidity, major axis length and Minor axis length are extracted by taking the features from the segmented leaf image. The mean, standard deviation, skewness and kurtosis for red, green and blue color features are extracted. Wiener filtering and canny edge detection are used to identify the vein features v1, v2, v3, v4 at different thresholds by calculating gray level value from the gray level histogram. The features are then exposed to Principal Component Analysis (PCA) for feature reduction and the resultant reduced feature set contains five shape features, six color features and three vein features. The features are then taken as input vectors for multi-layer neural network. Total 150 leaves from Columbia leaf image database are taken as samples. 150 leaves represent 10 different leaf species. The system was trained with 104 leaf images and was tested and validated using 23 leaf images each. The success rate of reduced feature set of leaf shapes and color was 84.7% and by combining leaf shape, color and vein features outputted a success rate of 95.3%.