Classification of weed seeds using image processing and pattern recognition methods

In order to identify weed seeds by machine vision approach, two methods of artificial neural network (ANN) and canonical discriminant analysis (CDA) were applied. The seeds of Amaranthus retroflexus, Amaranthus blitoides, Chenopodium album, Alyssum hirsutum, and Sinapis arvensis were collected and the images of these seeds were recorded. The obtained images were processed and then characteristics related to the shape of each seed were extracted. The extracted data were in in the raw and standardized forms. In addition, main shape characteristics of the seeds were identified by stepwise regression. The results showed that the accuracy of the ANN constructed from the raw and the standard data were 84.30% and 83.39%, respectively. Identification accuracy of ANN was 84.30% and 83.39% for the raw and standard data extracted from stepwise regression. The results of CDA method showed that the identification accuracy of total raw and standard data were 84.9% and 84.7% respectively. Identification accuracy of this method was 82.6% for both raw and standard data extracted from stepwise regression. The highest identification accuracy in both methods (more than 95%) was belonged to A. retroflexus and A. hirsutum. In addition, the seed identification accuracy of C. album in both methods was higher than 87%. This suggests that the use of shape features in pattern recognition models had reliable potential in identifying the seeds of these weeds.