An Efficient Insect Pest Classification Using Multiple Convolutional Neural Network Based Models

Introduction

In this project, we implement many convolutional neural network-based models on insect pests recognition task. Including: attention, feature pyramid network residual attention networks, and fine-grained models (MMAL-networks). We using an ensemble technique to combine our models to obtain the robust model. We test our proposed methods on 2 published datasets including IP102: A Large-Scale Benchmark Dataset for Insect Pest Recognition and D0. The experimental results show that combining these convolutional neural network-based models can better perform than the state-of-the-art methods on these two datasets. For instance, the highest accuracy we obtained on IP102 and D0 is 74.13% and 99.78%, respectively_

Our paper can be found at https://arxiv.org/abs/2107.12189 and under the submission to the journal Applied Intelligence.

Requirement

• python 3.7.10
• torch 1.7.1
• numpy 1.19.5
• matplotlib 3.2.2
• TensorboardX 2.0
• sklearn 0.22.2
• skimage 0.16.2
• imblearn 0.4.3
• barbar

Dataset

In this work, we using two published datasets:

• IP102 (proposed in https://openaccess.thecvf.com/content_CVPR_2019/papers/Wu_IP102_A_Large-Scale_Benchmark_Dataset_for_Insect_Pest_Recognition_CVPR_2019_paper.pdf). Contact the author to get this dataset.
• D0 (proposed in https://www.sciencedirect.com/science/article/abs/pii/S0168169916308833). Download link https://www.dlearningapp.com/web/DLFautoinsects.htm.

Usage

To preproduct the result:

• Download all requirement packages.
• Main working directory must be "...\Improving-Insect-Pest-Recognition-by-EnsemblingMultiple-Convolutional-Neural-Network-basedModels\code".
• Download dataset
• Prepare the training data using command:

py data_prepare.py -data IP102 -root ...\\ip102_v1.1-002

Note: Downloaded IP102's root folder directory must contained the archived .tar flie and .txt files of listing samples for training, validating,and testing. Created root folder for training will be placed in the same directory with this code.

py data_prepare.py -data D0 -root ...\\d0

Note: Downloaded D0's root folder must contained 40 archived .zip files. Created root folder for training will be placed in the same directory with this code.

• Training phase:
  • Training and Testing ResNet50

    python Trainmain.py -data IP102(or D0) -optim Adam -sch expdecay -l2 0.00001 -do 0.3 -predt True -mn resnet -lr 0.0001 -ep 100 -bz 64 -dv cuda

  • Training and Testing RAN

    python Trainmain.py -data IP102(or D0) -optim SGD -sch myScheduler -l2 0.0000 -do 0.0 -predt False -mn residual-attention -lr 0.1 -ep 100 -bz 32 -dv cuda

  • Training and Testing FPN

    python Trainmain.py -data IP102(or D0) -optim Adam -sch expdecay -l2 0.00001 -do 0.0 -predt True -mn fpn -lr 0.0001 -ep 100 -bz 32 -dv cuda

  • Training MMAL-Net

    python train.py -data IP102(or D0) -dv cuda

  • Testing MMAL-Net

    python test.py -data IP102(or D0) -dv cuda

• Ensemble all models with soft voting

  python implement_ensemble.py -data IP102(or D0) -dv cuda

Contact

• Nguyen Thanh Binh (University of Science Ho Chi Minh city, ngtbinh@hcmus.edu.vn)
• Ung Trung Hieu (ungtrunghieu99@gmail.com)
• Ung Quang Huy (ungquanghuy93@gmail.com)

References

 @misc{ung2021efficient,
      title={An Efficient Insect Pest Classification Using Multiple Convolutional Neural Network Based Models}, 
      author={Hieu T. Ung and Huy Q. Ung and Binh T. Nguyen},
      year={2021},
      eprint={2107.12189},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}