[Paper Review] Deep-Wide Learning Assistance for Insect Pest Classification

Nguyen Duc Toan, Huy Nguyen+3 more|arXiv (Cornell University)|Sep 16, 2024

Smart Agriculture and AI4 citations

TL;DR

This paper proposes DeWi, a novel one-stage deep-wide learning framework that enhances insect pest classification by jointly optimizing triplet margin loss for discriminative feature learning and Mixup-based data augmentation for generalization. DeWi achieves state-of-the-art performance with 76.44% accuracy on IP102 and 99.79% on D0, outperforming existing methods through a simple, efficient, and backbone-agnostic design.

ABSTRACT

Accurate insect pest recognition plays a critical role in agriculture. It is a challenging problem due to the intricate characteristics of insects. In this paper, we present DeWi, novel learning assistance for insect pest classification. With a one-stage and alternating training strategy, DeWi simultaneously improves several Convolutional Neural Networks in two perspectives: discrimination (by optimizing a triplet margin loss in a supervised training manner) and generalization (via data augmentation). From that, DeWi can learn discriminative and in-depth features of insect pests (deep) yet still generalize well to a large number of insect categories (wide). Experimental results show that DeWi achieves the highest performances on two insect pest classification benchmarks (76.44\% accuracy on the IP102 dataset and 99.79\% accuracy on the D0 dataset, respectively). In addition, extensive evaluations and ablation studies are conducted to thoroughly investigate our DeWi and demonstrate its superiority. Our source code is available at https://github.com/toannguyen1904/DeWi.

Motivation & Objective

• To address the challenges of fine-grained insect pest classification due to high intra-class variance and low inter-class variance.
• To improve model generalization across a wide range of pest species despite long-tailed data distributions and small-scale appearances.
• To develop a lightweight, efficient, and backbone-agnostic learning framework that enhances both feature discrimination and generalization in a single training stage.
• To outperform existing ensemble or complex architectures with a simpler, more practical solution suitable for real-world smart agriculture deployment.

Proposed method

• DeWi employs a one-stage training strategy that alternately optimizes triplet margin loss for discriminative feature learning and Mixup-based data augmentation for generalization.
• The triplet margin loss is applied in a supervised manner to enhance inter-class separation and intra-class compactness using hard mining.
• Mixup is used with α=1.0 to create uniform mixed samples, improving model robustness and generalization across diverse pest categories.
• The framework is compatible with various CNN backbones (e.g., ResNet-152), enabling broad applicability without architectural overhaul.
• The method avoids complex ensemble models or multi-branch networks, focusing instead on efficient joint optimization of discrimination and generalization.
• Ablation studies validate the effectiveness of the loss and augmentation components, including batch size and α hyperparameter sensitivity.

Experimental results

Research questions

• RQ1 Can a one-stage training strategy effectively balance deep feature learning and wide generalization in insect pest classification?
• RQ2 How does DeWi’s joint optimization of triplet margin loss and Mixup compare to self-supervised pretext-based pre-training in terms of performance and efficiency?
• RQ3 What is the optimal hyperparameter setting for Mixup (α) and batch size in the context of insect pest recognition?
• RQ4 How robust is DeWi to small batch sizes, and does it outperform traditional self-supervised pre-training under such constraints?
• RQ5 To what extent does DeWi improve performance across different CNN backbones and benchmark datasets?

Key findings

• DeWi achieves 76.44% accuracy, 69.46% mF1, and 65.07% GM on the IP102 dataset, significantly outperforming prior state-of-the-art methods.
• On the D0 dataset, DeWi achieves 99.79% accuracy, demonstrating exceptional performance on a highly fine-grained benchmark.
• The ablation study confirms that α=1.0 yields the best Mixup performance, with 76.44% accuracy, outperforming α=0.5 and α=2.0.
• DeWi maintains strong performance across different batch sizes (32, 24, 16, 8), showing robustness even with small batches.
• DeWi outperforms a self-supervised pretext-based setting by 13.36 pp in accuracy, 15.74 pp in mF1, and 20.26 pp in GM, highlighting the superiority of its one-stage approach.
• The method is effective across multiple CNN backbones and shows consistent improvements, validating its flexibility and generalization capability.