R-UNET: Improving performance of rumex detection using unet based selective regions

Abstract

In particular, Rumex weed detection is regarded as a crucial step in real-world data under many circumstances. The detection task suffers from several issues, such as overlapping weeds, occlusion, varying leaf colour distributions, leaf size and shape, and growth stage. Many machine learning techniques have been proposed to detect weeds in plants. These techniques suffer from locating weed with precise bounding boxes because they may contain multiple bounding boxes in a certain region.
Researchers have used the R-CNN based weed identification system, but it continues to have a low detection rate because of the issues mentioned above. In order to detect Rumex weeds under various conditions, particularly overlapping, occlusion, and size, as well as containing multiple bounding boxes, this paper is developed the R-CNN model by using UNet instead of the CNN model to become R-UNet. The proposed model is used due to its novelty of using a UNet classifier with selective regions which boosts the detection capabilities by extracting the most helpful features more effectively than the CNN network.
The proposed method uses Intersection over Union (IoU) to assess the detection rate using real-world data. We compare and benchmark the evaluation of the detection performance of this work with different models, including Single-Shot Detector (SSD), hybrid CNNs, AlexNet, and adapted NMS methods. The proposed model yields the highest IoU values compared with other methods.