
PeerJ Computer Science, Год журнала: 2024, Номер 10, С. e2518 - e2518
Опубликована: Ноя. 22, 2024
Extracting the essential features and learning appropriate patterns are two core character traits of a convolution neural network (CNN). Leveraging traits, this research proposes novel feature extraction framework code-named 'HierbaNetV1' that retrieves learns effective from an input image. Originality is brought by addressing problem varying-sized region interest (ROI) in image extracting using diversified filters. For every sample, 3,872 maps generated with multiple levels complexity. The proposed method integrates low-level high-level thus allowing model to learn intensive features. As follow-up research, crop-weed dataset termed 'SorghumWeedDataset_Classification' acquired created. This tested on HierbaNetV1 which compared against pre-trained models state-of-the-art (SOTA) architectures. Experimental results show outperforms other architectures accuracy 98.06%. An ablation study component analysis conducted demonstrate effectiveness HierbaNetV1. Validated benchmark weed datasets, also exhibits our suggested approach performs well terms generalization across wide variety crops weeds. To facilitate further weights implementation made accessible community GitHub. extend practicality, incorporated real-time application named HierbaApp assists farmers differentiating Future enhancements for outlined article currently underway.
Язык: Английский