Towards Accurate Disease Segmentation in Plant Images: A Comprehensive Dataset Creation and Network Evaluation

Komuravelli Prashanth, Jaladi Sri Harsha, Sivapuram Arun Kumar, Jaladi SrilekhaWACV 2024

TL;DR

This paper introduces a new, comprehensive dataset for segmenting Northern Leaf Blight (NLB) in maize plants, a crucial step for automated disease detection. They then evaluate state-of-the-art deep learning models, specifically Mask R-CNN and Cascaded Mask R-CNN with various backbones, demonstrating that the "ResNet Strikes Back-50" backbone and the cascaded architecture yield the best segmentation performance.

What Problem Does This Paper Address?

Agricultural productivity is significantly threatened by plant diseases, and early, accurate detection is key to mitigating yield losses. Traditional manual inspection methods are slow, subjective, and often miss early disease signs. While computer vision and deep learning offer promising automated solutions, their effectiveness hinges on high-quality, well-annotated datasets. Creating such datasets, especially with precise segmentation masks of diseased areas, is a labor-intensive challenge. This paper tackles the specific problem of segmenting Northern Leaf Blight (NLB) in maize plants by addressing the need for a comprehensive dataset and evaluating how different deep learning architectures perform on this task.

Methods

The authors employ two prominent deep learning architectures for instance segmentation:

  • Mask R-CNN: An extension of Faster R-CNN, it adds a mask prediction branch to simultaneously predict bounding boxes, class labels, and pixel-level segmentation masks for each object instance.
  • Cascaded Mask R-CNN: This architecture builds upon Mask R-CNN by introducing a cascaded approach. It refines object proposals through multiple stages, aiming to improve the accuracy and precision of segmentation masks, especially for intricate details.

These models are evaluated with several backbone architectures, which are responsible for feature extraction from the input images:

  • Swin-Tiny: A lightweight variant of the Swin Transformer, known for its efficiency and ability to capture hierarchical representations.
  • ResNet-18, ResNet-50: Standard convolutional neural network architectures with varying depths, providing a balance between performance and computational cost.
  • ResNet Strikes Back-50: An enhanced version of ResNet-50, incorporating modifications to potentially improve its feature representation power.
  • ResNet-50 + DCNv2: This combination integrates Deformable Convolutional Networks (DCNv2) with ResNet-50. DCNv2 allows the network to adaptively learn sampling locations within convolutional kernels, which is beneficial for handling irregular shapes and deformations common in disease spots.

Datasets

The primary contribution in terms of data is the creation of a new, comprehensive dataset for Northern Leaf Blight (NLB) disease segmentation in maize plants.

  • NLB Disease Segmentation Dataset: This dataset comprises 1000 high-resolution maize plant images. It includes detailed segmentation masks for NLB-affected regions, capturing variations in growth stages, lighting conditions, and disease severity. The dataset was created by first using an "Segment Anything" model for initial mask generation, followed by meticulous manual refinement by domain experts using the VGG Image Annotator (VIA) tool. The dataset was split into a training set (900 images) and a test set (100 images).

Key Insights and Ideas

  • Dataset Quality is Paramount: The paper strongly emphasizes that the availability of high-quality, well-annotated datasets is a critical bottleneck for developing accurate disease segmentation models. Their new dataset aims to alleviate this.
  • Cascaded Architectures Offer Refinement: The Cascaded Mask R-CNN consistently outperformed its single-stage Mask R-CNN counterpart across various backbones, suggesting that iterative refinement of predictions is highly beneficial for precise disease segmentation.
  • Backbone Matters Significantly: The choice of backbone architecture has a substantial impact on segmentation performance. Some backbones are better equipped to capture the fine-grained details of disease spots than others.
  • "ResNet Strikes Back-50" Shows Promise: This particular backbone architecture emerged as the top performer, indicating that its architectural enhancements are effective for NLB disease segmentation.
  • DCNv2 Aids in Irregular Shapes: The integration of DCNv2 with ResNet-50 showed improved results, highlighting its utility in handling the irregular and varied shapes of disease lesions.

Key Results

The study evaluated models using mean Average Precision (mAP) at IoU thresholds of 0.5 and 0.75.

  • Best Performing Model: The Mask R-CNN with the ResNet Strikes Back-50 backbone achieved the highest mAP score of 0.433 (at IoU 0.5) and 0.475 (at IoU 0.75).
  • Cascaded vs. Single-Stage: Cascaded Mask R-CNN generally achieved higher mAP scores than Mask R-CNN for most backbone combinations. For instance, with the ResNet-50 backbone, Cascaded Mask R-CNN achieved an mAP of 0.368, while Mask R-CNN achieved 0.356.
  • Swin-Tiny Performance: The Swin-Tiny backbone showed moderate performance, with Mask R-CNN achieving an mAP of 0.239 and Cascaded Mask R-CNN achieving 0.347.
  • ResNet-50 + DCNv2: This combination also showed strong results, with Cascaded Mask R-CNN achieving an mAP of 0.374 and Mask R-CNN achieving 0.375.

Limitations and Future Work

  • Limitations: The paper acknowledges that while their dataset is comprehensive, further expansion with more diverse plant species, diseases, and environmental conditions would enhance model generalizability. The current study focuses on a single disease (NLB) in a single crop (maize).
  • Future Work: The authors suggest exploring advanced techniques such as data augmentation strategies and ensemble methods to further boost accuracy and generalization. They also propose that their work can serve as a foundation for selecting appropriate models and backbones for future research and practical applications in plant disease segmentation.

Summary updated

BibTeX
@inproceedings{prashanth2024towards,
  author = {Komuravelli Prashanth and Jaladi Sri Harsha and Sivapuram Arun Kumar and Jaladi Srilekha},
  title = {{Towards Accurate Disease Segmentation in Plant Images: A Comprehensive Dataset Creation and Network Evaluation}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2024},
}
Figure