[Paper Review] Farm land weed detection with region-based deep convolutional neural networks

Mohammad Ibrahim Sarker, Hyongsuk Kim|arXiv (Cornell University)|Jun 5, 2019

Smart Agriculture and AI18 references4 citations

TL;DR

This paper proposes a region-based fully convolutional network (R-FCN) integrated with a fine-tuned ResNet-101 for accurate and efficient weed detection in farm land images. To address limited dataset size, the authors apply data augmentation and dropout to reduce overfitting, achieving superior detection performance on a small-scale weed dataset compared to standard CNN approaches.

ABSTRACT

Machine learning has become a major field of research in order to handle more and more complex image detection problems. Among the existing state-of-the-art CNN models, in this paper a region-based, fully convolutional network, for fast and accurate object detection has been proposed based on the experimental results. Among the region based networks, ResNet is regarded as the most recent CNN architecture which has obtained the best results at ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) in 2015. Deep residual networks (ResNets) can make the training process faster and attain more accuracy compared to their equivalent conventional neural networks. Being motivated with such unique attributes of ResNet, this paper evaluates the performance of fine-tuned ResNet for object classification of our weeds dataset. The dataset of farm land weeds detection is insufficient to train such deep CNN models. To overcome this shortcoming, we perform dropout techniques along with deep residual network for reducing over-fitting problem as well as applying data augmentation with the proposed ResNet to achieve a significant outperforming result from our weeds dataset. We achieved better object detection performance with Region-based Fully Convolutional Networks (R-FCN) technique which is latched with our proposed ResNet-101.

Motivation & Objective

• To develop a robust deep learning model for accurate detection of weeds in agricultural field images.
• To address the challenge of limited annotated weed image data in training deep convolutional neural networks.
• To improve generalization and reduce overfitting in small datasets using dropout and data augmentation techniques.
• To evaluate the performance of region-based fully convolutional networks (R-FCN) with residual networks for object detection in agricultural imagery.
• To achieve high detection accuracy and speed for real-time farm land weed monitoring applications.

Proposed method

• A region-based fully convolutional network (R-FCN) is employed to enable end-to-end object detection with spatial invariance.
• A pre-trained ResNet-101 architecture is fine-tuned on the farm land weed dataset to adapt features to the specific detection task.
• Dropout layers are applied during training to mitigate overfitting due to the small size of the weed dataset.
• Data augmentation techniques are used to artificially expand the training set and improve model robustness.
• The model is trained end-to-end using stochastic gradient descent with weight decay and momentum optimization.
• Feature maps from the final convolutional layer are used to generate region proposals and classify weeds with bounding boxes.

Experimental results

Research questions

• RQ1 Can a region-based fully convolutional network with a fine-tuned ResNet-101 achieve high accuracy in detecting weeds from farm land images?
• RQ2 How effective are dropout and data augmentation in reducing overfitting when training deep CNNs on small agricultural image datasets?
• RQ3 Does the integration of R-FCN with ResNet-101 outperform standard CNN architectures in weed detection tasks?
• RQ4 What is the trade-off between detection speed and accuracy when using deep residual networks in real-world farming applications?
• RQ5 Can transfer learning from ImageNet pre-trained models significantly improve performance on a limited weed detection dataset?

Key findings

• The proposed R-FCN with fine-tuned ResNet-101 achieved superior detection performance compared to standard CNN models on the small weed dataset.
• The use of data augmentation and dropout significantly reduced overfitting, improving model generalization on unseen farm land images.
• Fine-tuning the ResNet-101 backbone on the weed dataset led to faster convergence and higher accuracy than training from scratch.
• The region-based detection framework enabled precise localization of weeds with bounding boxes, suitable for targeted herbicide application.
• The model demonstrated robustness to variations in lighting, soil conditions, and weed appearance due to data augmentation.
• The approach achieved a notable balance between detection accuracy and inference speed, suitable for real-time field deployment.