Source-linked AI summary
Unsupervised Image Translation using Adversarial Networks for Improved Plant Disease Recognition
Haseeb Nazki, Sook Yoon, Alvaro Fuentes, Dong Sun Park
TL;DR
Plant disease recognition is hindered by costly data acquisition and severe class imbalance in limited datasets. The paper introduces AR-GAN-based synthetic augmentation for unsupervised image translation, reporting a 5.2% accuracy improvement versus 0.8% for classical augmentation under equal class distribution.
Problem
Plant disease recognition requires costly data collection and faces limited, severely imbalanced datasets, making early automated recognition challenging.
Method
AR-GAN combines adversarial, cycle-consistency, and activation reconstruction losses to generate synthetic plant disease images for dataset augmentation and balancing.
Results
+5.2% classification accuracy was achieved with generated synthetic samples, compared with +0.8% using classical augmentation in an equal class distribution environment.
Takeaways & Limitations
Synthetic augmentation using AR-GAN improved plant disease recognition performance in the studied data-deficient setting and produced visually convincing images with meaningful features.
Takeaways & Limitations
GAN training is unstable and can suffer from mode collapse, while unsupervised translation assumes meaningful correlation between unpaired domains.
Abstract
from arXiv · showhide
Acquisition of data in task-specific applications of machine learning like plant disease recognition is a costly endeavor owing to the requirements of professional human diligence and time constraints. In this paper, we present a simple pipeline that uses GANs in an unsupervised image translation environment to improve learning with respect to the data distribution in a plant disease dataset, reducing the partiality introduced by acute class imbalance and hence shifting the classification decision boundary towards better performance. The empirical analysis of our method is demonstrated on a limited dataset of 2789 tomato plant disease images, highly corrupted with an imbalance in the 9 disease categories. First, we extend the state of the art for the GAN-based image-to-image translation method by enhancing the perceptual quality of the generated images and preserving the semantics. We introduce AR-GAN, where in addition to the adversarial loss, our synthetic image generator optimizes on Activation Reconstruction loss (ARL) function that optimizes feature activations against the natural image. We present visually more compelling synthetic images in comparison to most prominent existing models and evaluate the performance of our GAN framework in terms of various datasets and metrics. Second, we evaluate the performance of a baseline convolutional neural network classifier for improved recognition using the resulting synthetic samples to augment our training set and compare it with the classical data augmentation scheme. We observe a significant improvement in classification accuracy (+5.2%) using generated synthetic samples as compared to (+0.8%) increase using classic augmentation in an equal class distribution environment.
1. Introduction
Plant disease recognition is hindered by costly data collection, limited labeled datasets, and class imbalance. The paper proposes AR-GAN-based synthetic augmentation combined with CNN recognition to improve data diversity and recognition performance.
- Motivation: Plant disease recognition is difficult because symptoms vary across plant parts, making early detection and treatment challenging.Automating recognition from images is motivated by the need to reduce crop damage and improve yield.
- Motivation: Small, expensive datasets and class imbalance limit supervised learning, while classical transformations provide relatively little diversity.Synthetic samples are proposed to introduce more variability and enrich training data.
- Contributions: AR-GAN extends GAN-based image translation by adding Activation Reconstruction loss to adversarial and cycle-consistency objectives.The added loss targets perceptual quality while preserving image structure and semantics.
- Contributions: The proposed pipeline combines unsupervised AR-GAN synthetic augmentation with a supervised CNN recognition system using real and generated training data.Generated samples are used to augment and balance the plant disease dataset.
2.1. Data Augmentation and Generation
Data augmentation addresses the need for more training examples, but classical image transformations provide limited variation. Generative models offer a way to create samples with greater diversity.
- Classical augmentation: Machine learning algorithms require large training datasets, motivating classical augmentation methods such as flips, crops, zooms, rotations, and translations.Elastic deformations have also been reported to improve augmentation in data space.
- Generative augmentation: GANs generate synthetic samples intended to share the characteristics of the training distribution.The generator maps random noise z to samples G(z; θ) approximating the data distribution p_d.
- Generative augmentation: GAN-based image translation is used because other generative models tend to produce blurrier images in practice.The paper therefore focuses its generation discussion on GANs.
2.2. Generative Adversarial Networks (GANs)
GANs learn synthetic data generation through competition between a generator and discriminator. Their effectiveness is accompanied by training instability and mode collapse concerns.
- GAN framework: GANs comprise a generator that produces synthetic data and a discriminator that estimates whether inputs are real or generated.The two networks are trained in opposition, with the generator attempting to fool the discriminator.
- GAN objective: The GAN objective combines the discriminator’s real-data and generated-data terms in a two-player optimization game.The generator seeks to produce samples that the discriminator classifies as real.
- GAN objective: The non-saturating generator loss maximizes log D(G(z)) to avoid zero-gradient learning when the discriminator rejects generated samples confidently.This objective preserves the gradient direction while preventing saturation.
- Challenges: GAN training is unstable, and mode collapse can cause generated points to concentrate in a common space.WGAN and WGAN-GP are discussed as approaches addressing distribution similarity and collapse-related problems.
2.3. Image-to-Image Translation
Image-to-image translation maps one representation of a scene to another, including transformations between grayscale and RGB images. Unsupervised approaches learn mappings between domains without corresponding paired images.
- Definition: Image-to-image translation converts one possible representation of a scene into another, such as grayscale images into RGB images.The paper distinguishes supervised translation, which has corresponding target-domain ground truth, from unsupervised translation.
- Unsupervised translation: Unsupervised translation uses two generators for opposite domain mappings and two discriminators that assess domain membership.Reconstruction or self-consistency losses preserve the input through the translation cycle.
2.4. Plant disease recognition
Plant disease recognition research spans shallow feature-based methods and deep learning approaches, including CNNs and transfer learning. Prior work demonstrates recognition across diverse crops, diseases, and imaging settings.
- Plant disease recognition methods include shallow approaches using hand-crafted features and deep learning approaches using image-based models.
- Prior tomato disease studies include thermal and stereo imaging, classic feature extraction with SVMs, greenhouse recognition, and transfer learning.
- Deep CNNs have been applied to identify 26 diseases across 14 crop species using more than 50,000 PlantVillage images.
2.5. Network Overview and Architecture
The proposed pipeline combines AR-GAN-based synthetic data augmentation with CNN recognition to address class imbalance and limited training data. AR-GAN extends cycle-consistent unpaired translation with activation reconstruction, while adversarial and reconstruction constraints regulate realism and correspondence.
- The pipeline generates synthetic samples with AR-GAN, adds them to real data, and trains a CNN recognition system on the combined dataset.
- AR-GAN: AR-GAN extends CycleGAN with an activation reconstruction module alongside adversarial and cycle-consistency losses.
- AR-GAN: Unpaired translation uses generators in both directions and discriminators for each domain, with cycle-consistency reconstructing the original input after translation.
- AR-GAN: Activation reconstruction compares feature representations of real and translated images to encourage perceptual realism and model stability.
- AR-GAN: The ARL weight λ trades off resemblance to the target domain against faithfulness to the input, and its suitable value depends on the domain.
- Architecture: The implementation uses ResNet-style generators with nine residual blocks, 70 × 70 PatchGAN discriminators, and a feature extractor initialized with ImageNet weights.
- Recognition system: A baseline ResNet50 classifier is used to compare classical and synthetic augmentation techniques.
3. Results
The study evaluates both the generative model and the recognition system across several datasets, including experiments on the tomato plant disease dataset. It compares AR-GAN with unpaired translation methods, studies the ARL term, and evaluates the complete recognition pipeline.
- The experiments evaluate AR-GAN and the recognition system on several datasets and compare the method with recent unpaired image-to-image translation approaches.
- The study includes analyses of the ARL term, variant comparisons, and evaluation of the proposed pipeline on the tomato plant disease dataset.
3.1. Datasets
The evaluation uses multiple datasets for generative-model assessment and a tomato plant disease dataset for recognition experiments. The tomato dataset contains images spanning nine disease classes and varied acquisition conditions, with ROI examples covering diseases and pests.
- The generative model is evaluated on Cityscapes, the tomato plant disease dataset, and datasets from prior image-to-image translation studies.
- The tomato dataset contains 2,789 plant images across 9 identifiable disease classes.
- Images were collected under varied illumination, season, temperature, humidity, location, infection status, plant size, infected region, and resolution conditions.
- ROI examples cover nutritional excess or deficiency, powdery mildew, gray mold, plague, canker, whitefly, leaf mold, low temperature, and miner.
3.2. Metrics
The paper evaluates its generative model with FCN, FID, and NIMA, while recognition performance is assessed using class precision and total accuracy.
- FCN evaluates generated-photo comprehensibility by comparing predicted semantic label maps with input ground-truth labels.The comparison uses mean Intersection-over-Union, per-class accuracy, and per-pixel accuracy.
- FID measures similarity between generated and real images, with smaller values indicating better quality.The paper notes that FID correlates with human judgment and captures similarity better than the Inception Score.
- NIMA estimates photographic aesthetic quality and perceptual relevance, with larger scores indicating better aesthetic quality.The reported NIMA value is averaged across test-set images.
- Class precision and total accuracy evaluate recognition performance, with class precision capturing prediction consistency for diseased-class images.
3.3. Results
AR-GAN was evaluated against image-translation variants and used to augment an imbalanced tomato disease dataset. Synthetic augmentation improved recognition more than classical augmentation, while AR-GAN produced higher-fidelity translations with fewer artifacts than CycleGAN.
- Comparing AR-GAN against other GAN models: AR-GAN was compared with supervised and unsupervised image-to-image translation methods using FCN, FID, and NIMA metrics across Cityscapes and tomato leaf data.The experiments also varied generator architectures and applied activation reconstruction loss in one or both translation directions.
- Comparing AR-GAN against CycleGAN variants: AR-GAN outperformed the evaluated architectural variants and produced translations with higher fidelity, greater aesthetic value, and fewer artifacts than CycleGAN.CycleGAN often confused foreground and background, whereas AR-GAN generated compositions more faithful to the input.
- Analysis of loss function: λ controls the trade-off between resemblance to the input and preservation of visual traits from the target domain.The suitable value depends on the two domains and the specific translation task.
- Analysis of classic and synthetic augmentation techniques: The RoI-extracted tomato dataset was heavily imbalanced: Leaf Mold and Whitefly exceeded half the data, while Low Temperature and Powdery Mildew each occupied less than 5%.The study used original, classically augmented, and synthetically augmented dataset instances to address data insufficiency and imbalance.
- Analysis of classic and synthetic augmentation techniques: 5.2%: synthetic augmentation increased recognition accuracy from 80.9% to 86.1%, compared with a 0.8% increase to 81.7% using classical augmentation.The baseline ResNet50 recognition CNN was evaluated on 200 unseen test samples per disease class.
- Analysis of classic and synthetic augmentation techniques: Synthetic augmentation increased true-positive rates across all disease classes and yielded longer decreases than increases in false positives relative to classical augmentation.These changes were associated with improved overall accuracy.
4. Discussion
The work combines AR-GAN-based synthetic augmentation with deep CNN classification to improve plant disease recognition from a relatively small dataset. AR-GAN adds Activation Reconstruction loss to improve generated-image perceptual quality, while synthetic samples improve classification accuracy over classic augmentation.
- AR-GAN combines adversarial and cycle-consistency objectives with Activation Reconstruction loss to measure feature activations against real images.
- The framework generates synthetic plant disease samples from a finite number of raw images and evaluates them across different datasets and metrics.
- The study tests synthetic samples as training-set augmentation for a baseline convolutional neural network classifier and compares them with classic augmentation.
- AR-GAN-generated images show meaningful features and visualizations that may be incorporated into other computer-aided algorithms.