WheatNet: A Lightweight Convolutional Neural Network for High-throughput Image-based Wheat Head Detection and Counting

1 Introduction

• Manual counting and narrow growth-cycle windows limit wheat-head observations, while image-based wheat analysis at scale remains a challenge.Breeding organizations record only a small subset of possible observations each season; wheat-head counts help estimate yield performance.Source: §1 · p. 2
• Image-based analytics for wheat were relatively untouched, and prior YOLOv4 work did not provide a complete, scalable, deployable framework.The authors identify minimal inference time and in-field deployability as requirements for high-throughput wheat-head detection and counting.Source: §1 · p. 2
• WheatNet is presented as robust to variation in image scale, orientation, size, lighting, and environmental conditions.Source: §1 · p. 3
• A lightweight encoder and decoder reduce learned parameters to support deployment on mobile devices for in-field use.Source: §1 · p. 3
• Parallel subnetworks perform density-based counting and localization simultaneously, which the authors say mutually strengthens prediction accuracy.Source: §1 · p. 3
• The framework targets real-time inference for yield estimation and high-throughput phenotyping, using point-level annotations that are less labor-intensive than box-level annotations.Source: §1 · p. 3

2 Methodology

• Point-level annotations mark head centers for both counting and localization, and are less labor-intensive and time-consuming than box-level annotations.The paper describes point annotations as especially practical for dense agricultural images where public annotated datasets are scarce.Source: §2 · p. 3
• Wheat-head localization and counting must handle variation in shape, color, orientation, scale, lighting, development stage, and overlaps.The method estimates head centers, following crowd-counting approaches for dense objects.Source: §2 · p. 3

2.1 Network Architecture

• The truncated MobileNetV2 uses its first four bottlenecks, merges their multi-scale feature maps, and sends the result to counting and localization branches.Feature maps are resized with deconvolution and nearest upsampling before concatenation; the last three bottlenecks are removed, which the authors report improves prediction accuracy while decreasing computation time.Source: §2.1 · p. 4, Figure 2 · p. 4
• The counting branch estimates a density map whose sum gives the head count, while the localization branch predicts head-center points.The counting branch uses dilated convolutions; the localization branch applies average pooling before selecting local peaks as predicted centers.Source: §2.1 · p. 5

2.2 Network Loss

• The network combines Euclidean loss for density-map estimation with pixel-wise Focal loss for localization, using L = Lden + β Lloc and β = 0.01.Focal loss down-weights well-classified examples to help address foreground-background imbalance; the experiments set γ = 2 and α = 0.25.Source: §2.2 · pp. 5–6

• Experiments were conducted in TensorFlow on an NVIDIA Tesla V100 GPU.Source: §3 · p. 6

3.1 Dataset

• The public dataset contains 3,373 RGB images from 10 locations across Europe, North America, Asia, and Australia, with varied field and camera conditions.Images were captured 1.8–3 m above the wheat heads; 20% were randomly selected for testing and the rest used for training.Source: §3.1 · p. 6

3.1.1 Ground Truth Density Maps

• Ground-truth density maps are formed by convolving head-point annotations with a Gaussian kernel, with the map sum matching the image’s total head count.The Gaussian standard deviation is estimated from the average distance to k-nearest neighboring head annotations.Source: §3.1.1 · p. 6

3.1.2 Ground Truth Localization Maps

• Ground-truth localization maps mark a small neighborhood around each annotated wheat head as positive using a 3 × 3 kernel.Source: §3.1.2 · p. 7

3.1.3 Data Augmentation

• Training augmentation uses image-pyramid scales from 0.4 to 1.0 in 0.2 increments, nine random 300 × 300 crops per scale, random flips, and Gaussian noise.Source: §3.1.3 · p. 7

• Model performance is evaluated with mean absolute error (MAE) and root mean squared error (RMSE).Source: §3.2 · p. 7

3.3 Training Details

• 97,128 augmented patches were randomly split into 90% training data and 10% validation data.Validation data monitored the training process.Source: §3.3 · p. 7
• Training ran for 130,000 iterations, with the learning rate decaying from 3 × 10^-4 to 1.5 × 10^-6.Weights used Xavier initialization; batch normalization was not used.Source: §3.3 · p. 7
• Optimization used mini-batches of 16 with SGD and the Adam optimizer.The passage describes SGD for optimization using Adam.Source: §3.3 · p. 7

3.4 Design of Experiments

• The comparison includes seven state-of-the-art models applicable to image-based object counting.Source: §3.4 · p. 7
• SCAR uses spatial- and channel-wise attention, while ASD combines parallel filter-size branches with adaptive weighting and SPN uses parallel dilated convolutions.These density-based models address image context or scale variation through different mechanisms.Source: §3.4 · pp. 7–8
• CSRNet pairs a truncated VGG feature extractor with dilated convolutions, while SaCNN combines feature maps from two VGG scales.Source: §3.4 · p. 8
• Faster R-CNN uses a region proposal network with a COCO-pretrained ResNet50 backbone, whereas SSD is a single-shot detector with a COCO-pretrained Inception-V2 backbone.Source: §3.4 · p. 8

3.5 Final Results

• WheatNet outperformed the other methods by varying extents on 675 test images; its counting branch outperformed its localization branch, and its localization branch surpassed all methods except Faster R-CNN.Table 3 compares counting performance; the proposed (Avg) result averages the two branch estimates.Source: §3.5 · p. 8, Table 3 · p. 8
• The authors calculate an overall prediction error percentage of approximately 0.087 from an MAE of 3.85 across 675 images.Source: §3.5 · p. 9
• WheatNet has the fewest parameters and requires fewer computation resources; its inference time is 0.54 s on an Intel Core-i9 CPU at 2.3 GHz.The method and other density-based methods use point-level annotations, unlike object detectors using box-level annotations.Source: §3.5 · p. 9
• Figure 3 presents input images, estimated density maps, estimated localization maps, and detected wheat images in successive rows.Its labels distinguish ground-truth counts from the counting-branch and localization-branch predictions.Source: Figure 3 · p. 9
• WheatNet's localization performance is competitive with Faster R-CNN in the visual comparison of predicted wheat-head centers.Figure 4 rows show input images, WheatNet detections, and Faster R-CNN detections.Source: §3.5 · p. 9, Figure 4 · p. 10

4 Analysis

• Pre-harvest yield estimation is motivated as a way to inform grain-management decisions, while manual head counting is labor-intensive and prone to human error.Source: §4.1 · p. 10
• The procedure averages heads per foot of row and kernels per head, assumes 22 kernels per head in practice, and uses row distance in inches.The passage gives a typical range of 18–26 kernels per head.Source: §4.1 · p. 10
• The yield formula estimates bushels per acre; its 0.48 factor incorporates kernel weight and unit conversions.The authors state that small differences in formula inputs can produce large differences in yield estimates.Source: §4.1 · pp. 10–11
• The proposed approach can perform hundreds of field replications for a more stable estimate, compared with the five locations Lyon and Klein suggest.Figure 5 illustrates the procedure with an estimated 42 heads and 12-inch row spacing.Source: §4.1 · p. 11, Figure 5 · p. 11

5 Conclusion

• A truncated MobileNetV2 encoder merges feature maps across scales, while parallel branches perform density-based counting and localization from point-level annotations.The authors report that the lightweight encoder-decoder reduces network parameters.Source: §5 · p. 11
• The paper also presents a wheat-yield estimation framework using traditional agronomic formulas.Source: §5 · p. 11
• WheatNet’s fast, compact model supports live in-field mobile use, with representative images intended to provide accurate and consistent yield estimates.The authors contrast this with manually traversing fields and repeatedly counting heads.Source: §5 · p. 11