Cott-ADNet: Lightweight Cotton Detection

• Cott-ADNet is a lightweight real-time object detector engineered for cotton boll and flower recognition under challenging field conditions.
• The model incorporates specialized modules such as ODConv, SPDConv, NGAM, and DRFSPPF to enhance feature extraction and robustness while maintaining low computational cost.
• It enables applications in automated harvesting, yield estimation, and phenotypic analysis by accurately detecting small, occluded targets amid cluttered backgrounds.

Searching arXiv for the specified paper and closely related entries to ground the article in the current literature. Search results confirm the cotton-detection paper "Cott-ADNet: Lightweight Real-Time Cotton Boll and Flower Detection Under Field Conditions" (Wang et al., 15 Sep 2025) and a separate, unrelated QR-code deblurring work using the name ADNet (Li et al., 14 Oct 2025). Cott-ADNet is a lightweight, real-time object detector for cotton boll and flower recognition under complex field conditions. It is built on YOLOv11n and is motivated by automated harvesting, yield estimation, and high-throughput phenotypic analysis, where detection must remain reliable despite small and dense targets, frequent occlusion, low-contrast or weak-feature objects, rotational variation, multi-scale variation, and cluttered field backgrounds. The model introduces task-driven convolutional refinement through ODConv and SPDConv, a NeLU-enhanced Global Attention Mechanism (NGAM), and a Dilated Receptive Field SPPF (DRFSPPF). On its main benchmark, it reports 91.543% Precision, 89.753% Recall, 90.639% F1-score, 93.285% mAP50, 71.296% mAP, and 7.5 GFLOPs (Wang et al., 15 Sep 2025).

1. Scope, problem setting, and nomenclature

Cott-ADNet addresses cotton boll and flower detection under field conditions, where recognition quality has direct implications for harvest timing, robotic picking, yield estimation, and breeding research. The paper frames cotton harvesting as still dominated by labor-intensive manual picking, with low efficiency and possible loss from missing the optimal harvest window. Within this setting, a detector must be accurate, fast, and sufficiently lightweight for edge platforms such as UGVs and UAVs (Wang et al., 15 Sep 2025).

The model is tailored to a specific visual regime. Cotton bolls and flowers are described as small and dense, frequently occluded, low-contrast or weak-feature objects, rotated and multi-scale, and embedded in cluttered scenes containing leaves, stems, shadows, and changing illumination. The paper’s central claim is that many prior cotton detection systems either lack robustness under real field conditions or are too computationally expensive for on-device deployment (Wang et al., 15 Sep 2025).

A common source of confusion is nomenclature. The term “ADNet” also appears in an unrelated QR code motion deblurring framework that is an adaptive dual-network system for decoding-oriented restoration (Li et al., 14 Oct 2025). In contrast, Cott-ADNet denotes the cotton boll and flower detector built on YOLOv11n for field deployment (Wang et al., 15 Sep 2025).

2. System architecture and design logic

Cott-ADNet retains the overall YOLOv11n framework but modifies the backbone and feature fusion path with specialized modules. The paper describes four principal changes: ODConv in shallow feature extraction, SPDConv in mid/high-level downsampling, NGAM in deep semantic fusion, and DRFSPPF in final pyramid aggregation. The stated objective is to improve spatial representation and robustness without a large FLOP increase (Wang et al., 15 Sep 2025).

The shallow layers use ODConv for directional and adaptive feature modeling. This choice is motivated by spatially varying cotton structures and by the need to model orientation changes, shape variations, and structural diversity. In the downsampling path, SPDConv is used at two downsampling points to reduce information loss for small targets. The deep feature stage inserts NGAM, which modifies GAM by replacing its internal activations with NeLU. The final aggregation stage replaces standard SPPF with DRFSPPF to expand receptive fields while maintaining low computational cost (Wang et al., 15 Sep 2025).

This configuration reflects a task-driven division of labor. ODConv and SPDConv emphasize preservation and discrimination of localized structure; NGAM targets weak and low-contrast feature learning; DRFSPPF supplies broader contextual modeling. A plausible implication is that the architecture is organized around the specific failure modes of cotton imagery rather than around a generic object-detection recipe.

3. Core modules and mathematical formulation

ODConv. The paper defines ODConv as

$$\mathrm{ODConv}(X) = \sum_{i=1}^{K} \alpha_i \cdot (W_i \ast X),$$

where $X$ is the input feature map, $W_i$ is the $i$-th learnable convolution kernel, $\alpha_i$ is a dynamic weight from a directional-attention mechanism, $K$ is the number of kernels, and $\ast$ denotes convolution. The stated interpretation is that ODConv adaptively mixes multiple kernels rather than relying on a fixed kernel response, improving sensitivity to orientation changes, shape variations, and structural diversity in cotton imagery (Wang et al., 15 Sep 2025).

SPDConv. To preserve small-object information, Cott-ADNet uses SPDConv instead of ordinary stride-based downsampling at two locations. The SPD transform is written as

$$X' = \mathrm{SPD}(X; scale) \in \mathbb{R}^{\frac{S}{scale} \times \frac{S}{scale} \times C\cdot scale^2}.$$

The subsequent convolution is given as

$X'' = \mathrm{Conv}_{1\times1,\;C_{\mathrm{out}(X')}.$

The paper notes that the typesetting in the second equation is partially garbled, but the intended meaning is that convolution compresses the channel-expanded representation after spatial rearrangement. The stated benefits are preservation of fine detail during downsampling, improved recall for small and partially occluded bolls, and better suitability to dense agricultural scenes than standard pooling or stride reduction (Wang et al., 15 Sep 2025).

NGAM and NeLU. NGAM is formed by replacing activations inside both the channel attention branch and spatial attention branch of GAM with NeLU. NeLU is defined as

$$\text{NeLU}(x)= \begin{cases} x, & x>0 \ -\dfrac{\alpha}{1+x^2}, & x\le 0 \end{cases}$$

with derivative

$X$0

and $X$1. The paper attributes three benefits to this replacement: better weak-feature representation for low-contrast cotton bolls and flowers, smoother attention interactions, and more stable training because negative inputs still carry non-zero gradient information. It further notes that NeLU avoids the exponential saturation issue of Sigmoid and is computationally light (Wang et al., 15 Sep 2025).

DRFSPPF. DRFSPPF extends SPPF through multi-scale pooling, large-kernel depthwise convolution, dilated depthwise convolution, residual recalibration, and output projection. For input $X$2, the paper specifies

$X$3

followed by

$X$4

then

$X$5

then residual recalibration

$X$6

and output projection

$X$7

The stated intuition is the combination of multi-scale pooling for pyramid context, large-kernel directional modeling for broader structure, dilated convolution for further receptive field expansion without many extra parameters, and residual recalibration to stabilize training and preserve original features (Wang et al., 15 Sep 2025).

4. Dataset construction and experimental protocol

The primary dataset contains 4,966 images drawn from public sources such as GitHub and Kaggle and from web-scraped images collected with a Python crawler. It includes four classes: FL for cotton flower, PB for partly opened cotton boll, DB for defected cotton boll, and FB for fully opened cotton boll. Images were standardized to 640 × 640, stored in jpg format, converted to YOLO format, and all labels were cross-validated by cotton breeding experts. The web-scraped images were augmented with random brightness adjustment, flipping, random masking, and Gaussian noise. The split is 3,982 training images, 494 validation images, and 490 test images, corresponding to an 8:1:1 partition. The paper states that roughly 78% of the images came from open-source repositories and 22% from web scraping (Wang et al., 15 Sep 2025).

An independent external validation set contains 1,216 images collected at the University of Georgia Gibbs Research Farm in Tifton, Georgia, USA, using a handheld mobile camera under varying field conditions including different lighting, occlusion, and clutter. This external set was reserved exclusively for evaluation rather than training, with the stated purpose of testing cross-domain generalization (Wang et al., 15 Sep 2025).

Training was conducted on a single NVIDIA V100 GPU with an 8-core CPU for preprocessing and augmentation. The reported settings are batch size 32, maximum epochs 800, AdamW as optimizer, initial learning rate 0.001, weight decay 0.0005, L2 regularization, and early stopping patience of 50. All models were trained with the same pipeline for fairness. Evaluation uses Precision, Recall, F1-score, mAP50, mAP averaged over IoU thresholds 0.5 to 0.95 in steps of 0.05, and GFLOPs, with

$X$8

Here, $X$9, $W_i$0, and $W_i$1 denote true positives, false positives, and false negatives, respectively (Wang et al., 15 Sep 2025).

5. Quantitative results and ablation evidence

The main benchmark results reported for Cott-ADNet are summarized below (Wang et al., 15 Sep 2025).

Metric	Value	Notes
Precision	91.543%	Abstract rounds to 91.5%
Recall	89.753%	Abstract rounds to 89.8%
F1-score	90.639%	Abstract rounds to 90.6%
mAP50	93.285%	Abstract rounds to 93.3%
mAP	71.296%	Abstract rounds to 71.3%
GFLOPs	7.5	Lightweight regime

The comparison set includes YOLOv8s and YOLOv8n, YOLOv9s and YOLOv9t, YOLOv10s and YOLOv10n, YOLOv11s and YOLOv11n, YOLOv12s and YOLOv12n, and RT-DETR-50. Among these models, the paper states that Cott-ADNet achieved the highest F1-score and the highest mAP50 among models below 10 GFLOPs while maintaining strong performance with only 7.5 GFLOPs. It also notes that YOLOv11s obtained the best mAP50 overall at 93.417% but required 21.6 GFLOPs, while RT-DETR-50 was much heavier at 130.5 GFLOPs (Wang et al., 15 Sep 2025).

The ablation study removes one component at a time and retrains the network. The full model yields P 91.543, R 89.753, F1 90.639, mAP50 93.285, mAP 71.296, and 7.5 GFLOPs. Removing ODConv gives P 92.334, R 87.097, F1 89.639, mAP50 92.571, mAP 72.527, and 7.6 GFLOPs. Removing SPDConv gives P 93.433, R 87.970, F1 90.619, mAP50 92.199, mAP 71.202, and 7.9 GFLOPs. Removing DRFSPPF gives P 90.486, R 86.211, F1 88.297, mAP50 90.925, mAP 71.068, and 7.3 GFLOPs. Removing NGAM gives P 90.232, R 87.399, F1 88.793, mAP50 92.045, mAP 72.068, and 6.2 GFLOPs. Removing NeLU gives P 91.736, R 88.885, F1 90.288, mAP50 93.077, mAP 72.664, and 7.5 GFLOPs. The paper interprets these results as showing that DRFSPPF and NGAM contribute the largest performance gains, SPDConv helps preserve small-object information, ODConv helps feature modeling with only slight cost, and NeLU improves training stability and balance without changing FLOPs (Wang et al., 15 Sep 2025).

6. Robustness, applications, and limitations

The paper explicitly states that Cott-ADNet maintains stable performance under multi-scale and rotational variations. It attributes this behavior to architectural choices: ODConv for directional variation, SPDConv for preserving small targets under scale reduction, DRFSPPF for expanded receptive field and multi-scale context, and NGAM for improved sensitivity to weak or blurred objects. This suggests that the model’s robustness is treated as a consequence of module specialization rather than as a separate post hoc property (Wang et al., 15 Sep 2025).

On the independent 1,216-image field set, the detector retained strong performance when objects were clearly visible. The remaining failure cases were mainly severe occlusion by foliage or branches, shadows and reflections, and confusion between partly opened and fully opened bolls. The paper therefore presents Cott-ADNet as generalizing well while also indicating that very heterogeneous field conditions remain challenging (Wang et al., 15 Sep 2025).

The stated practical uses are automated cotton harvesting, yield estimation, high-throughput phenotypic analysis, and edge deployment, with the paper explicitly suggesting future deployment on UAV and UGV platforms. It also mentions possible future extensions including multi-modal input such as thermal imagery, domain adaptation across regions, and end-to-end yield estimation pipelines. In this sense, Cott-ADNet is positioned not as a generic detector, but as a specialized agricultural detection model that combines lightweight convolutional design, NeLU-enhanced global attention, and dilated multi-scale receptive field expansion to meet the accuracy-efficiency constraints of in-field deployment (Wang et al., 15 Sep 2025).