A Fine-Tuned MobileNetv2 based Approach for Weed Classification

Weed Detection and classification plays a very crucial role in farming. Manually detecting weeds accurately and spraying herbicides exactly to the weed location is a challenge. Hence, there is a need for automation in weed detection. With the advancements in the field of computer vision, machine learning and deep learning, extensive research has been carried out in this field to accurately detect weeds and spray herbicides on the exact location of weeds so that the crops don’t get affected. It also aids in preventing health and environment hazards caused due to the unwanted spraying of herbicides. Hence, this paper proposes a fine-tuned MobileNetv2 based solution for automatic weed detection and classification in images. The proposed work transfer learns the rich feature extraction from the pre-trained MobileNetv2 architecture and fine-tunes the last layers of the model on DeepWeed dataset by considering a variety of optimizers and activation functions. The proposed work compares the performance of the fine-tuned model against several state-of-art algorithms on the DeepWeeds benchmark dataset. The results show that the proposed model outperforms other state-of-art methods.