SDE-DET: Detector for Pomelo Detection

• The paper demonstrates that SDE-DET integrates Star Block, Deformable Attention, and EMA into YOLOv8n to enhance detection accuracy in orchards.
• It achieves state-of-the-art performance on the STP-AgriData dataset with Precision of 0.883, Recall of 0.771, [email protected] of 0.838, and a compact 3.29M parameter model.
• The design effectively addresses challenges like multi-scale variation, occlusion, small-object detection, and background confusion in real-world orchard conditions.

SDE-DET is a lightweight, accuracy-oriented object detector for Shatian pomelo detection in complex orchard environments. It is described as an improved YOLOv8n detector designed for orchard scenes with multi-scale variation, occlusion from leaves and trunks, small objects, and background confusion. Its architecture combines three targeted modifications—Star Block, Deformable Attention, and Efficient Multi-Scale Attention (EMA)—and is evaluated on the custom STP-AgriData dataset. The reported final scores are 0.883 in Precision, 0.771 in Recall, 0.838 in mAP@0.5, 0.497 in [email protected]:0.95, and 0.823 in F1-score, and the paper states that SDE-DET achieved state-of-the-art performance on STP-AgriData (Hu et al., 24 Sep 2025).

1. Problem setting and detector identity

SDE-DET addresses Shatian pomelo detection in orchard imagery collected under practical field conditions. The paper identifies four central difficulties: multi-scale variation, because pomelos appear at very different sizes depending on camera distance; occlusion, because leaves and trunks frequently partially hide fruits; small-object detection, because distant pomelos occupy very few pixels; and background confusion, because pomelo color and texture can resemble leaves and branches (Hu et al., 24 Sep 2025).

A common misunderstanding is to treat SDE-DET as a wholly separate detection family. The paper instead describes it as an improved YOLOv8n detector. YOLOv8n is chosen as the baseline because it is lightweight and suitable for embedded and robotic deployment. SDE-DET modifies that baseline with three modules: Star Block for richer high-dimensional feature extraction with low overhead, Deformable Attention in the backbone for robustness under occlusion and geometric variation, and EMA in the neck for improved multi-scale fusion, especially for small objects (Hu et al., 24 Sep 2025).

This formulation places SDE-DET in the class of task-specific detector adaptations rather than generic, architecture-agnostic object detectors. The paper’s central claim is not that a generic detector suffices unchanged, but that orchard-specific failure modes can be addressed by targeted backbone and neck modifications.

2. STP-AgriData dataset and acquisition conditions

The paper introduces STP-AgriData, a dataset built specifically for Shatian pomelo detection. Its sources are field-collected images from Taipingshan Village, Ganzhou City, Jiangxi Province, China, together with publicly sourced online images intended to increase diversity and environment coverage. The field site is specified at coordinates $114^\circ51'45''$ E, $25^\circ20'52''$ N, with elevation 193.1 m and area about 49.45 mu (Hu et al., 24 Sep 2025).

The reported capture protocol is unusually detailed. Images were captured under natural light, mostly in the afternoon, on sunny days, during winter, and at the fruit stage from enlargement to maturity. Sampling selected mature trees aged 5–7 years, with tree height around 2 m. The camera was positioned on the outer south side of the canopy, with default aperture, shutter, and ISO settings. Resolution was $1920 \times 1080$, format was JPEG, camera distance to trees was 3–5 m, and camera height was 2 m (Hu et al., 24 Sep 2025).

After cleaning, the final dataset contains 317 images, with one damaged image removed. The split is 190 training images and 127 testing images, corresponding to a 6:4 ratio. To improve robustness and generalization, the 190 training images were augmented by brightness adjustment, contrast adjustment, denoising, grayscale conversion, horizontal flipping, and vertical flipping, producing an expanded training set of 1,330 images (Hu et al., 24 Sep 2025).

The dataset is integral to the method’s identity. The paper’s claims about state-of-the-art performance are specific to STP-AgriData, and the data collection protocol suggests that the model is optimized for orchard scenes characterized by moderate viewpoint variation, natural illumination, and dense canopy structure.

3. Network architecture and computational profile

SDE-DET is organized as a modified YOLOv8n pipeline with backbone, neck, and head components. The reported backbone sequence includes an initial convolution, Star Block, downsampling convolution stages, CSP/C2f-style feature extraction, SPPF, and Deformable Attention. The neck uses upsampling and feature fusion, with EMA inserted at multiple scales. The head remains a multi-scale detection head for different resolutions (Hu et al., 24 Sep 2025).

The model is reported as lightweight, with 3.29 M parameters, 32.4 GFLOPs, and a 6.69 MB model size. That compactness is central to the paper’s framing, because the target application is orchard automation, including automatic harvest robots (Hu et al., 24 Sep 2025).

The architectural logic is modular. Star Block is used early to reduce fine-grained feature loss, Deformable Attention is placed deep in the backbone to improve selectivity under occlusion, and EMA is inserted in the neck to strengthen multi-scale fusion. The paper presents these as complementary interventions rather than interchangeable plug-ins.

The design also reflects a specific view of failure propagation in small-object fruit detection. Early-stage information loss is treated as damaging to later feature quality; occlusion handling is assigned to attention-based spatial adaptation; and small-object recovery is treated as a neck-level fusion problem. This suggests a layered diagnosis of orchard detection errors rather than a single bottleneck.

4. Constituent modules

Star Block

The paper motivates Star Block by arguing that the original early convolutional stage in YOLOv8n creates a large dimensional gap between the input and the first feature map, causing loss of fine-grained information. Star Block, inherited from StarNet, is said to preserve and enrich features without a large computational increase. The manuscript states that channels are combined pairwise through element-wise multiplicative interactions, yielding a high-dimensional implicit feature mapping that behaves somewhat like a kernelized expansion (Hu et al., 24 Sep 2025).

In the reported redesign, the original initial convolutional block is replaced by a smaller 8-channel convolution, followed by a Star Block, then a 32-channel convolution. The intended effect is to preserve more original visual detail while increasing feature expressiveness with little extra parameter cost (Hu et al., 24 Sep 2025).

Deformable Attention

Deformable Attention is inserted in the backbone after SPPF and before the final backbone feature output. Its motivation is that standard attention uses fixed sampling positions and can struggle when fruit boundaries are partially hidden or spatially shifted. Deformable Attention learns adaptive offsets so that attention can focus on informative regions rather than fixed grid locations (Hu et al., 24 Sep 2025).

The paper gives the standard projection form

$Q = W_q x,\quad K = W_k x,\quad V = W_v x$

and describes bilinear interpolation at shifted sample points. It then applies standard multi-head attention with

$$h^{(m)} = a\!\left(q^{(m)} k^{(m)T}\right) v^{(m)}, \quad m=1,\dots,M,$$

followed by output concatenation. The stated benefits are improved occlusion robustness, adaptability to local spatial distortions, stronger emphasis on visible fruit parts, and greater robustness against cluttered orchard backgrounds (Hu et al., 24 Sep 2025).

Efficient Multi-Scale Attention (EMA)

EMA is introduced in the neck to address small-object detection and efficient multi-scale fusion. The paper describes EMA as preserving information across channels while reducing computational overhead, re-arranging channels into groups, learning different semantics from grouped sub-features, and aggregating cross-spatial information. If the channel tensor is split into groups,

$$X = [X_0, X_1, \dots, X_{G-1}], \qquad X_i \in \mathbb{R}^{\frac{C}{G}\times H \times W},$$

EMA applies 2D global average pooling

$O_c = \frac{1}{H \times W} \sum_i \sum_j X_c(i,j).$

The paper states that this encodes global context and long-range dependencies (Hu et al., 24 Sep 2025).

EMA is presented as especially important for distant pomelos that occupy small image regions and therefore lose discriminative detail easily. The claimed effect is improved small-object detection, better feature fusion across scales, pixel-level attention on high-level features, and greater robustness under clutter and partial visibility (Hu et al., 24 Sep 2025).

5. Evaluation protocol and reported performance

The paper evaluates SDE-DET under standard object-detection metrics. It explicitly uses

$\text{Precision} = \frac{TP}{TP + FP},$

$\text{Recall} = \frac{TP}{TP + FN},$

$AP = \int p(r)\,dr,$

$25^\circ20'52''$0

$25^\circ20'52''$1

and

$25^\circ20'52''$2

The experimental environment is Linux with AMD EPYC 7542 CPU, NVIDIA GeForce GTX 4090 GPU, PyTorch 2.4.0, CUDA 12.1, Python 3.11.5, batch size 8, input size $25^\circ20'52''$3, 100 epochs, and default learning rate and weight decay settings (Hu et al., 24 Sep 2025).

The comparison set includes YOLOv8n, YOLOv8s, YOLOv9c, YOLOv10n, YOLOv10s, Faster R-CNN, Cascade R-CNN, CenterNet, RTMDet-m, DDQ-4scale, DINO-4scale, and RT-DETR. The paper reports that SDE-DET improves [email protected] over YOLOv8n by 5.1%, over YOLOv8s by 2.9%, over YOLOv9c by 3.4%, over YOLOv10n by 7.3%, and over YOLOv10s by 5.9%. Against other mainstream detectors, reported [email protected] improvements are +25.7% over Faster R-CNN, +22.0% over Cascade R-CNN, +15.7% over CenterNet, +8.9% over RTMDet-m, +12.1% over DDQ-4scale, +9.8% over DINO-4scale, and +5.1% over RT-DETR (Hu et al., 24 Sep 2025).

Model	[email protected]	[email protected]:0.95
YOLOv8n	0.787	0.458
YOLOv8s	0.809	0.467
RT-DETR	0.787	0.459
SDE-DET	0.838	0.497

A separate comparison on parameter count reinforces the lightweight framing. The paper lists 40.34 M parameters for Faster R-CNN, 88.06 M for Cascade R-CNN, 32.24 M for CenterNet, 24.70 M for RTMDet-m, 48.03 M for DINO-4scale, 47.21 M for DDQ-4scale, 42.76 M for RT-DETR, and 3.29 M for SDE-DET (Hu et al., 24 Sep 2025).

These numbers support the paper’s dual claim: the detector is positioned as both compact and accurate. The reported advantage is not only absolute detection quality on STP-AgriData but also a favorable accuracy-to-size tradeoff.

6. Ablation, qualitative behavior, and limitations

The ablation study attributes the final performance to cumulative gains from the three added modules. Starting from baseline YOLOv8n, the paper reports 0.862 Precision, 0.700 Recall, 0.787 [email protected], 0.458 [email protected]:0.95, and 0.773 F1. Adding Deformable Attention yields 0.870, 0.716, 0.794, 0.456, and 0.786. Adding Star Block to that configuration yields 0.875, 0.738, 0.817, 0.487, and 0.800. Adding EMA produces the full SDE-DET result of 0.883, 0.771, 0.838, 0.497, and 0.823 (Hu et al., 24 Sep 2025).

Configuration	Precision	Recall	[email protected]	[email protected]:0.95	F1
Baseline YOLOv8n	0.862	0.700	0.787	0.458	0.773
+ Deformable Attention	0.870	0.716	0.794	0.456	0.786
+ Deformable Attention + Star Block	0.875	0.738	0.817	0.487	0.800
+ Deformable Attention + Star Block + EMA	0.883	0.771	0.838	0.497	0.823

The paper summarizes the baseline-to-final gains as +2.1% in Precision, +7.7% in Recall, +5.1% in [email protected], +3.9% in [email protected]:0.95, and +5.0% in F1. The qualitative analysis states that baseline YOLOv8n misses small distant fruits, suffers in strong lighting, makes false positives when leaves resemble pomelo, and struggles with dense overlapping fruit clusters. SDE-DET is reported to reduce these issues and to produce more confident detections. Grad-CAM heatmaps are said to indicate more accurate focus on fruit boundaries and color and edge cues (Hu et al., 24 Sep 2025).

The paper also states several limitations. Missed detections and false detections are not fully solved. Robustness could improve with more diverse training data. Future enhancements may include reinforcement learning, self-supervised learning, adaptation to other fruit types, and integration into real-world robotic workflows such as picking, grading, and yield prediction (Hu et al., 24 Sep 2025).

The significance of SDE-DET therefore lies less in a universal detection principle than in a concrete demonstration of domain specialization. A plausible implication is that orchard detection performance depends strongly on matching architectural bias to the visual failure modes of the target crop scene. In SDE-DET, Star Block targets early feature preservation, Deformable Attention targets partial visibility and geometric variation, and EMA targets multi-scale fusion for small fruits. The paper presents that combination as a reliable method for Shatian pomelo detection and as a foundation for the further development of automatic harvest robots (Hu et al., 24 Sep 2025).