d Weed Recognition in Agriculture Using Mask R-CNN

Thah Lieng Kang, - (2022) d Weed Recognition in Agriculture Using Mask R-CNN. [Final Year Project Report] (Unpublished)

	PDF Thah Lieng Kang - 24pages.pdf Download (381kB)
	PDF (Please get the password by email to repository@unimas.my , or call ext: 082-583914/3973/3933) Thah Lieng Kang- fulltext.pdf Restricted to Registered users only Download (8MB)

Abstract

Recent cooperation on deep learning has piqued the curiosity of those interested to utilise the techniques in agriculture. Weed management system is significant in agriculture that must be completed to improve crop production. The first step in weed management is to accurately classify the weeds and crops with an effective management strategy. Due to the enormous complexities in agricultural images, such as identical colour and texture, a deep neural network with pixel-wise grouping must be used to identify the plant type. The effectiveness of one of the most famous deep neural networks is examined in this paper to tackle the instance segmentation problems. Using field photos, Mask R-CNN is used to recognise weed plants (detection and classification). The dataset, which contains weeds and plants, is used to train a Mask R-CNN computer vision framework to classify and locate unique occurrences of weeds among plants. The dataset was trained on the MS COCO dataset, and the model was tailored to our classification purpose via transfer learning. Some well-reported problems in developing a suitable model are instance occlusion and the major resemblance between weeds and crops. Mask�RCNN is built on the FPN and the ResNet101 backbone. After the field images are tested on the pre-trained Mask R-CNN model, Mask R-CNN will give a class label and a bounding box offset for each weed and crop recognised. Moreover, the recognised weeds and crops will be given an object mask. Using the Mask R-CNN, the system can effectively perform instance segmentation on the images of weeds and crops with higher accuracy.

Actions (For repository members only: login required)

View Item