Source-linked AI summary
TasselNet: Counting maize tassels in the wild via local counts regression network
Hao Lu, Zhiguo Cao, Yang Xiao, Bohan Zhuang, Chunhua Shen
TL;DR
The paper addresses automated maize tassel counting under unconstrained field conditions, where environmental and plant-growth variations make accurate counting difficult. It proposes Tasselnet, a deep CNN using local-count regression, and reports state-of-the-art performance over previous baselines. The paper also constructs the MTC dataset and identifies dataset diversity and domain adaptation as future concerns.
Problem
Maize tassel counting remains mainly manual, while unconstrained field variations make accurate computer-vision counting challenging.
Method
Tasselnet is a deep CNN-based approach that regresses local tassel counts from image sub-regions instead of global or density-map outputs.
Results
Tasselnet achieves state-of-the-art performance and outperforms previous baseline approaches by large margins.
Takeaways & Limitations
Local-count regression is particularly suitable for counting problems whose ground-truth density maps cannot be precisely defined.
Takeaways & Limitations
The authors plan to enrich MTC because training-data diversity is important and identify domain adaptation for object counting as an open question.
Abstract
from arXiv · showhide
Accurately counting maize tassels is important for monitoring the growth status of maize plants. This tedious task, however, is still mainly done by manual efforts. In the context of modern plant phenotyping, automating this task is required to meet the need of large-scale analysis of genotype and phenotype. In recent years, computer vision technologies have experienced a significant breakthrough due to the emergence of large-scale datasets and increased computational resources. Naturally image-based approaches have also received much attention in plant-related studies. Yet a fact is that most image-based systems for plant phenotyping are deployed under controlled laboratory environment. When transferring the application scenario to unconstrained in-field conditions, intrinsic and extrinsic variations in the wild pose great challenges for accurate counting of maize tassels, which goes beyond the ability of conventional image processing techniques. This calls for further robust computer vision approaches to address in-field variations. This paper studies the in-field counting problem of maize tassels. To our knowledge, this is the first time that a plant-related counting problem is considered using computer vision technologies under unconstrained field-based environment.
Background
The paper frames maize tassel counting as an important but difficult in-field plant-phenotyping task. It introduces a local-count regression approach, a field dataset, and evaluations showing strong performance.
- Maize tassels are male flowers whose emergence marks reproductive stage, and their total number helps monitor growth status.
- Manual counting is inefficient and fallible, motivating automated image-based analysis for large-scale plant phenotyping.
- In-field counting is challenged by changing tassel appearance, cultivar differences, illumination, posture, occlusion, clutter, and image-quality variation.
- Existing tassel detection precision remains below 50%, while overlapping tassels expose limitations of detection-based counting.
- Tasselnet formulates the task as counting by regression and predicts local counts from densely sampled sub-images before merging them into a count map.
- The MTC dataset contains 361 images from 16 sequences across four Chinese experimental fields, with manually assigned dot annotations and tassel counts ranging from 0 to around 100.
- Evaluations on MTC report that Tasselnet outperforms state-of-the-art methods and substantially reduces counting errors.
Experimental fields and imaging devices
The study uses time-series imagery collected from four experimental fields across China with dedicated field imaging devices. Images were captured hourly from 9:00 to 16:00 from elevated vertical views.
- Sixteen independent time-series image sequences were collected from four experimental fields across China between 2010 and 2015.
- The dataset includes six maize cultivars: Jundan No.20, Nongda No.108, Wuyue No.3, Zhengdan No.32, Jidan No.20, and Tianlong No.9.
- Images were captured hourly from 9:00 to 16:00 using a fixed 16mm focal length and a five-meters-height vertical view, except four meters for Gucheng sequences.
- The imaging devices transmitted captured images to a remote server for user access.
Maize tassels counting dataset
The MTC dataset supports in-field maize tassel counting under varied conditions, while Tasselnet uses local-count regression with CNN architectures and regression losses.
- Dataset: Manual dot annotations provide both image counts and information about the spatial distribution of tassels.Each tassel receives one dot annotation, following a standard object-counting annotation paradigm.
- Local counts regression: Tasselnet regresses counts within small image patches rather than global counts or dense density maps.Local regression is intended to model local characteristics, avoid dense per-pixel learning, and provide many training patches.
- Prediction: During prediction, sliding-window local counts are aggregated into a full-resolution count map and normalized by pixel coverage.Each pixel is divided by the number of sub-images contributing a prediction there.
- Regression target: Ground-truth density maps are generated by Gaussian smoothing of dot annotations, but local counts integrated from those maps serve as the regression target.The density-map sum can be fractional near image boundaries because part of a Gaussian may fall outside the image.
- Network and loss design: Three CNN capacities and three regression losses are evaluated, including LeNet-like, AlexNet-like, VGG-VD16-Net-like, ℓ1, ℓ2, and Huber choices.The evaluated CNNs contain approximately 1.4×10^5, 2.5×10^5, and 2.4×10^6 parameters, respectively.
Evaluation metric
The paper evaluates counting performance with MAE and MSE, then examines how architecture, training-sample volume, loss, Gaussian smoothing, and sub-image size affect results.
- MAE measures estimate accuracy, while MSE assesses estimate robustness; lower values indicate better counting performance.
- Number of training samples: MAE decreases from 9.5 to 6.5 as the number of training sub-images increases.Performance is almost identical for sampling strides sr = 8 and sr = 4, suggesting moderate training volume can capture field variation well.
- Gaussian kernel parameter: Counting performance varies little across Gaussian kernel parameters σ = 4, 8, and 12, although the optimal value differs by test sequence.The paper recommends choosing σ by observing training-set Gaussian smoothing responses relative to the median tassel size.
- Sub-image sizes: The optimal performance does not correlate consistently with sub-image size across r = 16, 32, 64, and 96.The authors relate this variation to sequence-specific tassel sizes and recommend a relatively small, but not too small, sub-image size.
Comparison with the state of the art
Tasselnet is compared with segmentation, detection, global regression, global density regression, and local density regression baselines on the MTC dataset. It performs best overall while motivating local counts regression for variable-sized, occluded tassels.
- The comparison includes JointSeg, mTASSEL, GlobalReg, DensityReg, and Counting-CNN as segmentation, detection, global-regression, and density-regression baselines.
- Tasselnet outperforms the baseline approaches in 7 out of 8 test sequences and achieves the overall best counting performance.Its MAE and MSE are significantly lower than those of the competing methods.
- JointSeg and mTASSEL perform poorly, indicating that simple color-cue segmentation and standard object detection are inadequate for this in-field counting problem.The paper links detection difficulty to overlapping tassels and non-maximum suppression.
- Global regression can achieve counting performance comparable to mTASSEL despite using only global image features rather than bounding-box-level annotations.This supports formulating maize tassel counting as a counting-by-regression task.
- The authors recommend regression for heavily occluded objects, local counts regression for dramatically varying object sizes, small sub-images, moderately complex deep models, and initially trying ℓ1 loss.
- Local counts regression is presented as suitable when tassel density maps are difficult to define precisely because tassel sizes vary.The paper reports that Tasselnet achieves state-of-the-art performance and large-margin improvements over previous baselines.
Competing interests
The authors declare that they have no competing interests.
- The authors declare that they have no competing interests.
Availability of data and materials
The paper states that the MTC dataset and supporting materials will be made available online.
- The MTC dataset and other supporting materials will be made available online.
- The work was supported by Chinese meteorological public-welfare and National Natural Science Foundation grants.
Author’s contributions
HL led the study’s core research, experimentation, analysis, and manuscript drafting, while other authors supervised, supported experimental design and training, and contributed to writing.
- HL proposed local counts regression for maize tassel counting, implemented the technical pipeline, conducted experiments, analysed results, and drafted the manuscript.
- ZG, YX, and CS co-supervised the study and contributed to manuscript writing.
- BZ helped design the experiments and provided technical support for efficient model training.