Improving the Accuracy of Lettuce and Weed Classification Based on MobileNetV2 Features Through Segmentation

Abstract

Automating the separation of commodity crops and weeds is a major challenge in the implementation of precision agriculture . The presence of complex backgrounds such as soil, rocks, and shadows often degrades the performance of feature extraction in computer vision classification models. This study proposes an image preprocessing approach using the GrabCut segmentation method to extract key crop objects cleanly before performing Deep Learning- based feature extraction . Representative features from the image are extracted using the lightweight and efficient MobileNetV2 architecture. Next, classification is performed by comparing three Machine Learning algorithms , namely Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Random Forest (RF). Testing is carried out on two data scenarios, namely the original dataset ( Original ) and the segmented dataset ( GrabCut ). The experimental results show that the use of original images produces an accuracy of 98.89% for all three classification models. However, after being integrated with GrabCut segmentation, the accuracy of all three models increases significantly to 100.00%. These results prove that GrabCut-based segmentation effectively eliminates background noise information , thereby improving the generalization capabilities of classification models perfectly on edge computing devices .