Adaptive Non-Maximum Suppression for improving performance of Rumex detection

Abstract

A crucial post-processing stage in numerous object detection methods is Non-Maximum Suppression (NMS). The key idea of this technique is to rank the detected bounding boxes according to their scores. Subsequently, selecting the bounding box with the maximum score represents the one best-fitted to the object and suppresses the remaining significant boxes. Conventional NMS suffers from locating objects with accurate bounding boxes as there are multiple boxes in a certain region. This issue reduces the detection performance of automated weed applications in the real world. Weed detection methods based on Region-Convolutional Neural Network (R-CNN) frameworks remain suffer from a lack of detection rate due to overlapping and occlusion leaves issues. This paper presents an Ensemble-Region Convolutional Neural Networks (E-RCNN) model of three state-of-the-art networks to detect Rumex obtusifolius L. (R. obtu.) weeds under various conditions, especially overlapping. The proposed E-RCNN model is used due to its novelty of using ensemble classifiers with the combination of three extractors at its backbone. Adaptive Non-Maximum Suppression (ANMS) is proposed with the Region Proposal Network (RPN) to enhance the detection performance of overlapping and occluded objects by overcoming the drawbacks of conventional Non-Maximum Suppression (NMS). A hybrid model of three CNN extractor networks is used as the backbone in the classification stage. Thus, integrating three networks into one robust model increases the recognition capability by extracting additional useful features more efficiently than those from an individual network. For detection, RPN is used to generate multi-proposed boxes, whereas ANMS is used to select the best box that has a high score rate to match the target object. Our proposed model has trained and tested two standard benchmarking datasets of Rumex weeds under real-world data. The proposed model tested each dataset separately to evaluate the detection rate in terms of Intersection over Union (IoU). For comparing the evaluation of the detection rate, AlexNet, Single-Shot Detector (SSD), DetectNet and Faster R-CNN with conventional NMS models are used to compare the results.