Weed Detection in Agriculture using End-to-End Object Detection with Transformers in ResNet-50

Weed infestations significantly expose agricultural productivity to danger, leading to substantial yield losses in large-scale farming operations. Accurate identification and targeted treatment of various weed species are critical for effective weed control in contemporary farming practices. This research explores an innovative approach for weed detection utilizing the DETR (End-to-End Object Detection) model with the ResNet-50 backbone in a multi-label scenario, leveraging a custom-built dataset containing annotations for four distinct weed species present in 3956 images. This study diverges from prior methodologies and focuses on comprehensive weed detection using advanced deep learning architectures. Unlike previous approaches, the investigation specifically assesses the DETR-ResNet-50 model's efficacy in the precise identification and localization of multiple weed species within a single image. The research methodology encompasses annotations for multiple weed species in single images, covering diverse scenarios encountered in practical agricultural settings. Although primarily utilizing loss values for model evaluation, additional metrics such as mAP (mean average precision), IoU (intersection over union), precision, and recall were employed to assess the model's performance. The study concludes that the DETR-ResNet-50 model showcases promising potential for effective weed detection, signifying its viability for practical deployment in agricultural settings.

I. Introduction

Crop management stands as the cornerstone of successful agricultural practices, exerting a direct influence on the productivity and sustainability of crops. In this domain, the enduring challenge of weed infestation emerges as a crucial factor that profoundly affects crop management and agricultural yield. Weeds, in their quest for essential resources like sunlight, water, and nutrients, not only vie with crops but also introduce adverse effects that culminate in reduced crop yields and substantial economic losses within farming industries. The effective mitigation of weeds plays a central role in ensuring optimal crop growth and yield. Precise identification and management of a myriad of weed species prove indispensable for establishing robust weed control strategies. [1] Traditional methods employed for weed detection and classification have often been simplistic, lacking the capability to navigate the intricate dynamics inherent in multi-species weed populations prevalent across diverse agricultural settings. In this context, this research endeavor takes a significant leap forward by exploring modern deep learning architectures tailored for weed detection. Focused on the DETR (End-to-End Object Detection) model featuring the ResNet-50 backbone, this study [3] seeks to revolutionize weed detection methodologies. Unlike previous approaches that concentrated solely on individual weed seedling classification, this investigation pivots towards a more comprehensive framework.