Source-linked AI summary
Leaf Counting with Deep Convolutional and Deconvolutional Networks
Shubhra Aich, Ian Stavness
TL;DR
The paper addresses automated rosette-leaf counting from RGB images, motivated by the cost and limitations of manual plant phenotyping. It combines whole-plant segmentation with convolutional regression and reports a generalized framework evaluated across LCC-2017 datasets, while performance is constrained by scarce training data for some plant types.
Problem
The paper seeks robust rosette-leaf counting across plant species and image conditions, a task relevant to plant health and costly manual phenotyping.
Method
A deconvolutional SegNet-style network segments the whole plant, and a convolutional regression network estimates leaf count from RGB plus binary segmentation inputs.
Results
The framework learns leaf counts across five LCC-2017 test datasets using data-driven architectures and simple augmentation despite the dataset’s small training size.
Takeaways & Limitations
Direct regression with whole-plant segmentation provides a generalized alternative to individual-leaf instance segmentation for the evaluated datasets.
Takeaways & Limitations
Performance is poor for A3 and mixed A5 because only 27 Tobacco training images are available compared with 783 Arabidopsis images.
Abstract
from arXiv · showhide
In this paper, we investigate the problem of counting rosette leaves from an RGB image, an important task in plant phenotyping. We propose a data-driven approach for this task generalized over different plant species and imaging setups. To accomplish this task, we use state-of-the-art deep learning architectures: a deconvolutional network for initial segmentation and a convolutional network for leaf counting. Evaluation is performed on the leaf counting challenge dataset at CVPPP-2017. Despite the small number of training samples in this dataset, as compared to typical deep learning image sets, we obtain satisfactory performance on segmenting leaves from the background as a whole and counting the number of leaves using simple data augmentation strategies. Comparative analysis is provided against methods evaluated on the previous competition datasets. Our framework achieves mean and standard deviation of absolute count difference of 1.62 and 2.30 averaged over all five test datasets.
1. Introduction
The paper targets automated rosette-leaf counting as a plant-health indicator and seeks a robust framework that generalizes across species and imaging conditions. It uses data-driven segmentation and regression to support practical plant phenotyping.
- Motivation: Manual plant phenotyping is slow, tedious, expensive, and can introduce measurement bias through sparse random sampling.Automating visual-trait estimation could increase production speed and reduce costs.
- Problem: The target task is estimating the number of leaves on a rosette-stage plant, an indicator of plant health.
- Objective: The framework aims to generalize across plant species and image-data quality for use by plant breeders.
- Problem formulation: Leaf counting is posed as nonlinear regression, estimating counts directly without segmenting individual leaf instances.The paper states that this can be more efficient and requires simpler ground-truth annotation than instance segmentation.
- Framework: A single generalized network is evaluated across five LCC-2017 subsets, combining whole-plant segmentation with convolutional leaf counting.
- Intended use: The work seeks a useful generalized system for studying leaf development and, eventually, crop emergence in the field.
2. Related Work
Prior leaf-counting methods span competition-era segmentation, regression, and deep-learning approaches. Their differences center on feature design, instance-level processing, and reliance on ground-truth segmentations.
- LSC-2014: CVPPP-2014 methods used color and pixel features for foreground segmentation, followed by morphology, watershed transformation, and heuristic merging.The reviewed limitation is reliance on simple pixel features without deep contextual information.
- LSC-2014: The LSC-2014 winner modeled plant and background regions with 3D Lab-color histograms and inferred test pixels non-parametrically.
- LCC-2015: GLC introduced nonlinear regression for leaf counting using log-polar transformations, patch features, K-means encoding, max-pooling, and support vector regression.
- LCC-2015: GLC used ground-truth plant segmentations during training and testing, leaving robustness to noisy or imperfect segmentations unclear.
- Later methods: Later methods adopted deep learning, including recurrent instance segmentation with convolutional LSTM processing and deconvolutional architectures.
- Comparison: RIS and related instance-segmentation approaches require instance-level ground truth and are therefore not directly comparable with the paper’s approach.
3. Our Approach
The framework first segments the whole plant with a SegNet-style deconvolutional network, then estimates leaf count from combined segmentation and RGB inputs. Its design emphasizes semi-global context and reduced parameterization for a small dataset.
- Overall framework: The pipeline follows a segmentation-then-counting workflow using task-specific convolutional architectures trained without explicit plant-species knowledge.The two modules are trained separately but are coupled through the binary segmentation output.
- 3.1. Segmentation: Semantic segmentation separates plant or foreground pixels from background and produces class-probability maps through convolutional and deconvolutional subnetworks.
- 3.1. Segmentation: SegNet omits fully connected layers because semi-global context is considered sufficient for leaf-region segmentation, reducing trainable parameters and computational complexity.
- 3.1. Segmentation: The segmentation network compresses feature maps 32 times through five 2×2 pooling operations before deconvolutional reconstruction.
- 3.2. Counting: The counting network receives both the original RGB image and a binary segmentation mask rather than either input alone.This four-channel representation is called SRGB.
- 3.2. Counting: SRGB is intended to recover missed plant regions and reject false detections while reducing background-feature bias.
- 3.2. Counting: The regression network draws on VGG-style convolutional depth but uses unpadded convolutions to reduce parameters and limit overfitting on several hundred images.
- 3.2. Counting: The counting architecture estimates leaf number from SRGB channels using convolution, local response normalization, ReLU, and 2×2 max-pooling blocks.
4. Experiments
The experiments evaluate the framework on the LCC-2017 dataset using segmentation, counting comparisons, and ablation analyses. Results show strong generalized performance overall, but limited training data constrains performance on Tobacco images and mixed-source testing.
- Dataset: The LCC-2017 dataset provides RGB plant images with binary plant-background masks, leaf-center images, and ground-truth leaf counts across source-specific directories.Training includes 27 Tobacco images in A3 and 783 Arabidopsis images across the remaining directories; testing adds the wild A5 directory.
- Training and implementation: The segmentation network was trained from scratch with extensive augmentation, while the counting network used resized whole images and 36 augmented samples per original image.The counting network produced 29,160 training instances and was trained for 40 epochs; segmentation augmentation generated about 0.8M samples from 810 originals.
- Evaluation: The evaluation assessed plant-pixel precision and recall, compared the method with prior approaches, and included an ablation study of the segmentation input channel.The ablation trained counting networks with and without the segmentation channel.
- Evaluation: Segmentation results were considered sufficient for regression, although performance was lower on A3 because its training set contained only 27 Tobacco images.The paper reports semi-global context as adequate for whole-plant segmentation used by the regression network.
- Evaluation: The framework achieves lower CountDiff and AbsCountDiff than GLC on directories A1 and A2, indicating less count bias and better average performance there.CountDiff reflects over- or underestimation bias, while AbsCountDiff measures average absolute counting error.
- Evaluation: The method compares favorably with prior LSC-2014 and LCC-2015 approaches overall, but fails on A3 and shows degraded AbsCountDiff on mixed Arabidopsis-Tobacco directory A5.The authors attribute these failures to insufficient Tobacco training data for the deep regression architecture.
5. Conclusion and Future Work
The paper presents a generalized RGB leaf-counting framework that segments plants before counting, while acknowledging that segmentation’s necessity remains unresolved. Future work targets equivalent performance with simpler nonrecurrent networks.
- The framework counts leaves directly from RGB images without instance segmentation and generalizes across plant species and surroundings.
- Plant segmentation precedes counting because the authors assume a foreground channel helps the regression model extract features from the plant region.
- Whether deep networks require segmentation before counting remains an open question.
- Future work seeks performance equivalent to instance-segmentation architectures using simpler, easier-to-train nonrecurrent networks.