Modular-based classification system for weed classification using mixture of features

摘要

Spot weeding which could theoretically automate the application of herbicide can only be developed with the advancement of weed image processing recognition. A modular-based classifier is proposed in this work whereby neural network classifiers are developed to recognise the presence of a single type of weeds using a combination of various feature types. To demonstrate the concept, two types of young weeds were used which are the broadleaf weeds and narrow leaf weeds. The weeds cannot be distinguished by shape analysis alone as some specimens are overlapping with another. Hence, classifier for each individual species(species classifier) are developed by analysing/training with a several type of features such as co-occurrence matrix, Haralick features, shape analysis and histogram. Results indicate that relatively high recognition rate was acquired with selected features after feature selection search process. The recognition rate recorded using selected features are 98.8% for narrow leaf weeds and 100% for broad leaf weeds despite a high network size - 40 hidden neurons for broadleaf and 70 hidden neurons for narrow leaf weeds.