A hybrid CNN–SVM classifier for weed recognition in winter rape field

BACKGROUND: Weed recognition is key for automatic weeding, which is a challenging problem. Weed recognition is mainly based on different features of crop images. The extracted image features mainly include color, texture, shape, etc. The designed features depend on manual work, which is blind to some extent. Meanwhile these features have poor generalization performance on a sample set. The final discrimination results tend to have a greater difference. The current study proposed a deep convolutional neural network (CNN) with support vector machine (SVM) classifier which aims to improve the classification accuracy of winter rape seeding and weeds in fields. RESULTS: The VGG network model was adopted, which received a true color image (224 × 224 pixels) of rape/weed as the input. The proposed VGG-SVM model was able to identify rape/weeds with average accuracies of 99.2% in the training procedures and 92.1% in the test procedures, respectively. A comparative experiment was conducted using the proposed VGG-SVM model and five other methods. The proposed VGG-SVM model obtained a higher classification accuracy, greater robustness and real time. CONCLUSIONS: The VGG-SVM weed classification model proposed in this study is effective. The model can be further applied to the recognition of multi-sample mixed crop images in fields.