Feature Refinement Method Based on the Two-Stage Detection Framework for Similar Pest Detection in the Field

Hongbo Chen, Rujing Wang, Jianming Du, Tianjiao Chen, Haiyun Liu, Jie Zhang, Rui Li, Guotao ZhouInsects 2023 · DOI

TL;DR

This paper introduces a novel feature refinement method within a two-stage object detection framework to tackle the challenging problem of distinguishing visually similar pests in complex agricultural field environments. By enhancing feature representation, adaptively fusing multi-scale features, and separating classification and localization tasks, their approach achieves state-of-the-art performance on a newly created dataset of similar pests.

What Problem Does This Paper Address?

Accurate and timely pest detection is crucial for effective Integrated Pest Management (IPM) and ensuring food safety. However, identifying pests in real-world agricultural settings is fraught with difficulties. These include the inherent similarity in appearance (size, shape, color) among different pest species, especially in their larval stages, and the chaotic nature of field environments (varying lighting, occlusions, complex backgrounds). Traditional methods struggle with these challenges, leading to misidentification, which can result in the use of ineffective pesticides, crop damage, and environmental pollution. Existing deep learning approaches, while promising, often fall short when dealing with highly similar pests in unconstrained field conditions.

Methods

The authors propose a feature refinement method built upon a two-stage object detection framework, specifically adapting the Cascade RCNN architecture. Their approach introduces three key novel modules:

  • Context Feature Enhancement Module (CFEM): This module aims to enrich the feature representation by capturing multi-scale contextual information. It uses atrous spatial pyramid pooling (ASPP) to extract context at different receptive field sizes and then integrates this contextual information back into the original features via residual connections. This helps the model better understand the semantic meaning of pest regions within their surroundings.
  • RoI Feature Fusion Module (RFFM): To address the issue of suboptimal feature selection from a single scale, RFFM adaptively fuses features from multiple scales. It uses RoIAlign to extract features from different pyramid levels and then employs an attention mechanism to learn adaptive weights for fusing these features. This ensures that both detailed information from high-resolution maps and semantic information from low-resolution maps are effectively utilized.
  • Task Separation Module (TSM): Recognizing that classification and localization tasks might benefit from different feature representations, TSM decouples these two branches. Specifically, it uses features enhanced by CFEM for localization (which benefits from detailed information) and features fused by RFFM from the FPN for classification (which benefits from semantic information). This allows each task head to be optimized independently.

Datasets

The primary dataset used in this research is SimilarPest5, a newly constructed dataset specifically designed for detecting similar pests in field environments. It comprises 5177 images of five similar lepidopteran pest species: Spodoptera frugiperda, Mythimna separata, Ostrinia furnacalis, Helicoverpa armigera, and Spodoptera litura. The images were collected from real corn and soybean fields across China between 2020 and 2022, capturing diverse conditions. The authors also compared their method against existing datasets and benchmarks, noting that many existing datasets have less similar pest species or are collected under more controlled conditions.

Key Insights and Ideas

  • Context is King for Similar Pests: The CFEM highlights that simply extracting features isn't enough; understanding the pest's context within the scene is crucial for distinguishing visually similar species.
  • Adaptive Fusion Beats Fixed Fusion: The RFFM demonstrates that adaptively weighting and fusing features from multiple scales is more effective than relying on a single scale or a fixed fusion strategy, especially when dealing with varying pest sizes and appearances.
  • Task Specialization Boosts Performance: Separating the classification and localization tasks, and feeding them distinct, optimized feature sets (via TSM), leads to improved overall detection accuracy. This suggests that a one-size-fits-all feature representation might not be optimal for all sub-tasks within object detection.
  • A Dedicated Dataset Matters: The creation of SimilarPest5 underscores the need for specialized benchmarks to push the boundaries of research in niche areas like similar pest detection.

Key Results

The proposed feature refinement method, integrated into the Cascade RCNN framework, achieved a mean average precision (mAP) of 72.7% on the SimilarPest5 dataset. This represents a significant improvement over other state-of-the-art object detection methods, outperforming the best-performing algorithm by 1.2%. The method also showed strong performance in detecting medium and large-scale pests, achieving higher recall rates. Ablation studies confirmed the effectiveness of each proposed module (CFEM, RFFM, TSM), with the full model showing a 3.1% mAP improvement over the baseline Cascade RCNN.

Limitations and Future Work

The authors acknowledge that the proposed method, while effective, introduces additional modules, leading to increased computational complexity and a lower FPS compared to simpler models. They also note that the SimilarPest5 dataset, while diverse, could be expanded with more pest species and crop types.

Future work directions include:

  • Collecting and constructing larger datasets of field pests from various crops.
  • Extending the method to detect other types of pests (e.g., rice pests).
  • Further analyzing pest similarity based on morphological, ecological, and statistical characteristics.
  • Developing lighter model architectures to reduce computational complexity and improve efficiency for deployment on terminal devices.

Summary updated

Abstract

Efficient pest identification and control is critical for ensuring food safety. Therefore, automatic detection of pests has high practical value for Integrated Pest Management (IPM). However, complex field environments and the similarity in appearance among pests can pose a significant challenge to the accurate identification of pests. In this paper, a feature refinement method designed for similar pest detection in the field based on the two-stage detection framework is proposed. Firstly, we designed a context feature enhancement module to enhance the feature expression ability of the network for different pests. Secondly, the adaptive feature fusion network was proposed to avoid the suboptimal problem of feature selection on a single scale. Finally, we designed a novel task separation network with different fusion features constructed for the classification task and the localization task. Our method was evaluated on the proposed dataset of similar pests named SimilarPest5 and achieved a mean average precision (mAP) of 72.7%, which was better than other advanced object detection methods.
BibTeX
@inproceedings{chen2023feature,
  author = {Hongbo Chen and Rujing Wang and Jianming Du and Tianjiao Chen and Haiyun Liu and Jie Zhang and Rui Li and Guotao Zhou},
  title = {{Feature Refinement Method Based on the Two-Stage Detection Framework for Similar Pest Detection in the Field}},
  booktitle = {Insects},
  year = {2023},
  doi = {10.3390/insects14100819},
}
Figure