PaperlayerView exact revision on arXiv

Source-linked AI summary

High performing ensemble of convolutional neural networks for insect pest image detection

Loris Nanni, Alessandro Manfe, Gianluca Maguolo, Alessandra Lumini, Sheryl Brahnam

arXiv:2108.12539v1cs.CV

TL;DR

Automated insect identification from trap images addresses a costly, expertise-intensive task. The paper evaluates fused CNNs with new DGrad-based Adam variants on Deng and IP102, reporting state-of-the-art results on both datasets.

  • Problem

    Identifying and counting trapped insects is costly and time-consuming, and visual conditions such as small size, species resemblance, occlusion, and illumination complicate recognition.

    Source: §1 · p. 1

  • Method

    The paper fuses five CNN topologies trained with varied augmentation or Adam optimization, including two proposed DGrad variants that scale the learning rate.

    Source: §1 · pp. 2–3

  • Results

    The best ensemble achieved state-of-the-art results on both Deng and IP102, and the paper reports that HERE outperformed tested approaches on each dataset.

    Source: §5.3 · pp. 15, 17, Conclusion · p. 20

  • Takeaways & Limitations

    The proposed CNN ensembles and Adam variants produced top results on the two evaluated pest datasets, with additional medical-imaging tests also reported.

    Source: §1 · p. 3, Conclusion · p. 20

    Further evaluation on medical imaging datasets is needed to confirm the strength and robustness of the new Adam variants.

    Source: §5.3 · p. 18

Abstract

from arXiv · show

Pest infestation is a major cause of crop damage and lost revenues worldwide. Automatic identification of invasive insects would greatly speedup the identification of pests and expedite their removal. In this paper, we generate ensembles of CNNs based on different topologies (ResNet50, GoogleNet, ShuffleNet, MobileNetv2, and DenseNet201) altered by random selection from a simple set of data augmentation methods or optimized with different Adam variants for pest identification. Two new Adam algorithms for deep network optimization based on DGrad are proposed that introduce a scaling factor in the learning rate. Sets of the five CNNs that vary in either data augmentation or the type of Adam optimization were trained on both the Deng (SMALL) and the large IP102 pest data sets. Ensembles were compared and evaluated using three performance indicators. The best performing ensemble, which combined the CNNs using the different augmentation methods and the two new Adam variants proposed here, achieved state of the art on both insect data sets: 95.52% on Deng and 73.46% on IP102, a score on Deng that competed with human expert classifications. Additional tests were performed on data sets for medical imagery classification that further validated the robustness and power of the proposed Adam optimization variants. All MATLAB source code is available at https://github.com/LorisNanni/.

1 Introduction

Pest identification from trap images is costly and difficult because insects can be small, similar, damaged, or obscured. The paper proposes CNN classification with new DGrad-based Adam variants and evaluates fused classifiers on two pest benchmarks.

1 Introduction

  • Trap-image identification is time-consuming and requires taxonomic expertise, while species resemblance, trap discoloration, debris, clumping, occlusion, and illumination complicate automated recognition.Complex crop and ground backgrounds remain a challenge for identification.Source: §1 · pp. 1–2
  • The introduction situates the work amid expanding use of CNNs in pest classification and prior efforts to improve Adam, including methods that adjust parameter step sizes using gradient changes.The cited literature describes diffGrad and DGrad among gradient-difference-based approaches.Source: §1 · pp. 2–3
  • CNN training motivates optimization improvements because Adam tracks averages of past gradients and squared gradients, while the paper proposes two new DGrad variants for pest classification.Exp adds a scaling factor; ExpLR adds a step when calculating the final learning rate.Source: §1 · pp. 2–3
  • The study compares fused CNN classifiers trained with different Adam variants on Deng and IP102, and reports that its best method achieves state of the art on both datasets.Additional medical-image tests assess the proposed optimizer variants.Source: §1 · pp. 2–3

2 Related Work on Insect classification

Prior pest-classification research spans small handcrafted-feature datasets and increasingly large CNN benchmarks. Earlier ensembles and model comparisons provide reference results on Deng and IP102 for the present study.

2 Related Work on Insect classification

  • Earlier pest-image datasets were small and often supported handcrafted-feature pipelines, including SIFT, SURF, or HOG descriptors paired with SVMs.Source: §2 · p. 3
  • IP102 advanced deep-learning pest classification with over 75,000 images across 102 classes, while Deng contains 563 samples in ten classes.IP102 includes 19,000 images annotated with bounding boxes.Source: §2 · pp. 3–4
  • Earlier work compared CNNs across multiple pest datasets, using simple reflection, rotation, translation, and scaling augmentation to help prevent overfitting.The cited comparison evaluated pretrained ImageNet models on NBAIR, Xie1, and Xie2 datasets.Source: §2 · p. 4
  • A prior fused-CNN approach using saliency-preprocessed images reported 92.43% accuracy on Deng and 61.93% on IP102.Source: §2 · p. 4
  • GAEnsemble reported 95.16% accuracy on Deng and 67.13% on IP102, establishing a prior state-of-the-art reference in the cited work.Source: §2 · p. 5

3 Convolutional Neural Networks (CNNs)

The study uses five CNN architectures with differing designs and computational profiles. ShuffleNet and MobileNetv2 are included in part for speed, efficiency, and mobile-device suitability.

3 Convolutional Neural Networks (CNNs)

  • CNNs combine convolutional and pooling feature-extraction layers with fully connected classification layers and a softmax mapping to class labels.Convolution extracts region features, while pooling reduces feature dimensionality.Source: §3 · p. 5

3.1 ResNet50 (RN)

  • ResNet50 has 48 convolutional layers, max and average pooling, and more than 23 million trainable parameters; skip connections help it train many layers by preventing vanishing gradients.Source: §3.1 · p. 5

3.2 GoogLeNet (GL)

  • GoogleNet has 22 layers, or 27 including pooling layers, and four million trainable parameters; inception layers select among filter sizes within each block.Different filter sizes can handle different image patterns.Source: §3.2 · p. 5

3.3 ShuffleNet (SN)

  • ShuffleNet uses group convolutions and channel shuffling to reduce parameters, computation time, and the number of multiplications for mobile devices.Source: §3 · p. 5, §3.3 · p. 6

3.4 MobileNetv2 (MN)

  • MobileNetv2 uses depthwise separable convolutions to reduce memory requirements and inverted skip connections to reduce operations for mobile applications.Source: §3 · p. 5, §3.4 · p. 6

3.5 DenseNet201 (DN)

  • DenseNet201 has 201 layers and 20 million trainable parameters, concatenating layer outputs rather than using ResNet’s additive merging to increase gradient flow and feature reuse.Source: §3.5 · p. 6

4 Adam Optimization

The section motivates robust CNN optimization because deep architectures face vanishing gradients and difficult parameter searches, then outlines Adam and its variants, including the newly introduced Exp and ExpLR. These variants alter learning-rate behavior using gradient-related terms.

4 Adam Optimization

  • Exp and ExpLR are introduced as two new Adam variants in this section.Both build on DGrad, with Exp adding a learning-rate scaling factor and ExpLR adding a multiplicative sigmoid term.Source: §4 · p. 7, §4.6 · p. 9, §4.7 · p. 10

4.1. Adam

  • Adam computes adaptive parameter learning rates using gradients and exponential moving averages of first and second gradient moments.The moment decay rates are typically 0.9 and 0.999; bias correction addresses the small early averages caused by zero initialization.Source: §4.1 · p. 7

4.2. AMSGrad

  • AMSGrad replaces Adam’s exponential average of squared gradients with the maximum of previously squared gradients when updating parameters.The remaining parameter update is the same as Adam’s, apart from replacing that term.Source: §4.2 · p. 8

4.3. diffGrad

  • diffGrad uses the difference between consecutive gradients to set parameter learning rates, seeking larger steps for faster gradient change and smaller steps for slower change.Its learning-rate factor is the sigmoid of the gradient difference.Source: §4.3 · p. 8

4.4 DGrad

  • DGrad bases its gradient-change term on differences in absolute gradients and moving averages of element-wise squared gradients.It normalizes the term, multiplies it by 4 to increase the output range, and otherwise retains diffGrad’s final update.Source: §4.4 · pp. 8–9

4.5 Cos

  • Cos adds a cyclic learning rate to DGrad, multiplying its learning-rate factor by the cyclic rate.The cycle period is set to 30 steps, and the rate is 9 × 10^-4 when its value is zero.Source: §4.5 · p. 9

4.6 Exp (new)

  • Exp adds a scaling factor k to DGrad’s learning-rate term; higher k makes the curve steeper and more concentrated.The section tests k = 4 and defines the element-wise product and exponential used in its calculation.Source: §4.6 · pp. 9–10, Figure 2 · p. 10

4.7 ExpLR (new)

  • ExpLR adds a multiplicative sigmoid term to Exp’s final learning rate, using twice the normalized DGrad term as its input.The sigmoid normalizes the learning-rate output.Source: §4.7 · pp. 10–11, Figure 3 · p. 11

5 Experimental Results

Experiments compare CNN ensembles across the Deng and IP102 pest benchmarks under different training and augmentation choices. HERE is reported as state of the art on both, with additional tests evaluating the Adam variants and ensembles on medical and virus data.

5.1 Data sets

  • Deng contains 563 images in ten unbalanced pest classes; the reported protocol trains on 200 images and tests on 363, with results averaged across five splits.The split uses 20 randomly extracted training images per class.Source: §5.1 · pp. 11–12
  • IP102 contains 75,222 images across 102 pest classes, with class sizes from 71 to 5,740 and a 45,095-image training set plus a 22,169-image test set.Its hierarchical taxonomy and images depicting different insect development stages add to the classification challenge.Source: §5.1 · p. 12

5.2 CNN Settings

  • The authors use simple reflection and random scaling augmentation for the small Deng data set, but report that augmentation was unnecessary for IP102.CNNs were fine-tuned with a replaced output layer sized to the data set's class count and trained using cross-entropy; SGD runs also used randomly varied augmentation each epoch.Source: §5.2 · pp. 12–13, §5.3 · p. 13

5.3 Results

  • HERE is reported as state of the art on Deng and IP102; on Deng, the new Adam approaches outperform other Adam methods and SGD, and the approach comes close to the best human experts.The results are reported in the Deng accuracy table, and the human-expert comparison is described separately.Source: §5.3 · pp. 14–16, Table 2 · p. 14
  • HERE outperforms the other tested approaches on IP102 and is reported as state of the art for both F-score and G-means.These additional metrics are weighted averages of one-versus-all results, weighted by each test class's fraction of samples.Source: §5.3 · p. 17, Table 4 · p. 17, Table 5 · p. 17
  • On LAR, Exp+ExpLR achieves an F-measure of 97.04, exceeding the state-of-the-art results reported in and.Across the tested medical data sets, Exp and ExpLR are identified as the best-performing approaches, and their sum-rule fusion outperforms both base approaches and SGD.Source: §5.3 · pp. 18–19, Table 6 · p. 19
  • On the Kylberg Virus object-scale benchmark, the fusion of Exp and ExpLR achieves 92.53% accuracy, compared with 91.93 for Exp and 91.67 for ExpLR.The paper reports state-of-the-art performance on the object-scale benchmark; comparison with is not possible because its fixed-scale data set is unavailable.Source: §5.3 · pp. 19–20, Table 7 · p. 20

Conclusion

The paper combines CNN topologies and Adam optimization methods, including two proposed DGrad variants, and reports top results against the literature on pest and medical-imaging benchmarks. Future work will apply active learning to select new web images for labeling.

Conclusion

  • The combined CNN ensembles and proposed Adam variants achieve top results against the literature on both pest data sets and medical-imaging benchmarks.The paper identifies the two new Adam algorithms as DGrad-based variants that introduce a learning-rate scaling factor.Source: Conclusion · p. 20
  • Future work will use active learning to download many web images and select a small number of new images for labeling.The stated aim is to select images that might improve classification performance.Source: Conclusion · p. 20
  • The authors make all MATLAB source code available through GitHub.Source: Conclusion · p. 20
Loading 2108.12539v1…