PROTECT YOUR DNA WITH QUANTUM TECHNOLOGY
Orgo-Life the new way to the future Advertising by AdpathwayAgricultural engineers have developed an enhanced deep learning vision model designed to reliably identify green peppers obscured by dense foliage in commercial greenhouses, overcoming long-standing computer vision hurdles caused by identical fruit and leaf colorations. Cultivated under protected agricultural structures, green peppers share an almost indistinguishable green color gamut, surface texture, and local reflectance with their surrounding vegetative canopies. This severe visual uniformity, widely known as green-on-green background interference, typically impairs standard automated counting, harvest forecasting, and robotic yield management systems. To address these optical challenges, researchers led by Zhihao Zhang, Xue Qin, and Wei Li introduced a specialized neural framework termed FMDD-YOLO, integrating wavelet-based frequency analysis and deformable spatial convolution into the baseline YOLOv11m architecture to separate subtle pepper boundaries from surrounding plant foliage.
The foundational challenge of visual sensing in protected horticulture stems from the delicate physical and optical characteristics of pepper crops. Green peppers frequently grow in tight clusters where leaves, branching vines, and neighboring fruits overlap, generating fragmented contours, irregular shading, and mutual occlusion. In addition, peppers naturally exhibit slender, curved, and twisted morphologies that defy rigid geometric templates, while variations in camera distance across sprawling greenhouse rows generate dramatic shifts in target scales. Traditional machine vision techniques relying on handcrafted color thresholds, edge detectors, or shallow texture descriptors systematically break down under variable greenhouse illumination, direct sunlight glare, and complex foliage patterns. Although modern one-stage deep learning detectors provide real-time inference speeds, baseline models commonly suffer from high false-positive rates on leafy textures, miss small distant fruits, and produce misaligned bounding boxes when fruit margins blend seamlessly into background leaves.
To evaluate their architecture under realistic agricultural operating conditions, the research team assembled a dedicated visual dataset captured at a protected pepper cultivation facility in Kaiyang County, Guiyang, Guizhou Province, China. Collected between October 15 and November 10, 2024, during the peak fruit maturity period, the image repository encompassed diverse ambient lighting conditions, including direct sunlight, soft diffuse illumination transmitted through structural greenhouse plastic films, dense canopy shadows, and low-light environments. Using smartphone rear cameras stationed at standard close-range agricultural inspection distances between 30 and 80 centimeters, the team captured high-resolution images across frontal, lateral, and oblique vantage points. Rigorous screening eliminated blurred or duplicated frames, yielding 1,604 valid source images containing thousands of individual fruit instances distributed across varied cluster configurations, fruit sizes, and foliage occlusion levels.
Dataset integrity was maintained through rigorous independent dual-annotation protocols and structured data partitioning. Two experienced annotators independently labeled bounding boxes using visible contours and inferable spatial extents, classifying foliage occlusion into mild (under 30 percent covered), moderate (30 to 60 percent covered), and severe (over 60 percent covered) categories, while excluding ambiguous targets lacking sufficient visible area. An initial agreement rate of 89.2 percent was achieved based on an intersection-over-union threshold of at least 0.80, with all remaining discrepancies resolved through structured joint review. The original images were subsequently partitioned into training, validation, and testing sets following a strict 7:1:2 ratio, grouping images by individual plants, acquisition rows, and dates to eliminate potential visual data leakage. Data augmentation operations, including Gaussian noise injection, brightness adjustments, rotation, cropping, and flipping, were confined exclusively to the training split, leaving the independent test set entirely untouched.
The resulting FMDD-YOLO framework incorporates four complementary structural innovations mapped directly to the baseline YOLOv11m backbone, feature aggregation neck, and prediction head. In the feature extraction backbone, the standard bottleneck units of the C3k2 module were replaced with C3k2_WTConv, a module that incorporates wavelet convolutions. Standard spatial convolutions process image grids uniformly and often miss subtle high-frequency edge transitions between identically colored leaves and fruits. The WTConv mechanism applies a two-dimensional discrete wavelet transform to decompose input feature maps into four distinct frequency subbands: a low-frequency approximation that preserves broad object continuity and three directional high-frequency subbands that isolate sharp local boundaries. Operating in parallel with a spatial depthwise convolution pathway, the module learns optimal channel-wise frequency weights before applying an inverse wavelet transform, giving the network heightened sensitivity to subtle fruit contours embedded within leafy backgrounds.
Following spatial pyramid pooling in the deeper semantic stages of the network, the researchers introduced a multi-scale cascaded attention module, designated MSCAM, to suppress persistent non-fruit background noise. Deep convolutional features possess expansive receptive fields, yet their responses frequently blur fruit contours with neighboring stems, vines, and ground shadows. MSCAM combats this blending through an ordered three-tier pipeline: a feature enhancement module that uses parallel multi-scale and dilated depthwise convolutions to capture directional contours and wider spatial context, followed by a channel attention block that suppresses leaf-dominated channel responses, and a spatial attention block that concentrates activation peaks onto visible fruit bodies. By structurally enhancing intermediate features before computing channel and spatial attention weights, MSCAM ensures that valid fragments of partially occluded peppers are retained and amplified rather than prematurely filtered out.
To resolve feature distortion during cross-scale aggregation in the network neck, FMDD-YOLO substitutes standard nearest-neighbor or bilinear upsampling with a lightweight, content-aware dynamic upsampling module known as DySample. Conventional static interpolation relies on fixed geometric grids that inevitably copy surrounding foliage data into high-resolution feature maps when enlarging low-resolution representations. DySample addresses this limitation by generating point-sampling offsets dynamically conditioned on local feature content, steering the sampling points toward valid, unoccluded fruit areas and discriminative edges. This data-driven resampling prevents boundary blurring across feature pyramid levels, preserves the visual details of small and distant fruits, and ensures precise spatial alignment between high-level semantic abstractions and low-level boundary representations.
At the final stage of the network, the investigators designed DynamicDCNv4Head, replacing the baseline detection head with an adaptive mechanism that dynamically refines prediction features before passing them to the classification and bounding-box regression layers. Standard prediction heads apply uniform convolutions and static activation functions regardless of target orientation or background complexity. DynamicDCNv4Head incorporates deformable convolution version 4 to adjust spatial receptive fields to the irregular, curved shapes of hanging peppers, while a shared ScaleWeight generator dynamically balances contributions between adjacent feature pyramid levels based on object scale. Furthermore, a dynamic rectified linear unit, DyReLU, modulates channel responses based on multi-scale contextual inputs, strengthening target activations and dampening specular highlights. By confining these dynamic modifications to pre-prediction feature refinement, the architecture enhances localization without altering the fundamental YOLOv11 regression pipeline.
Controlled experimental evaluations demonstrated substantial quantitative gains across all primary detection metrics. Evaluated across five independent random training seeds on the self-built greenhouse test set, FMDD-YOLO attained a mean precision of 98.6 ± 0.2 percent, a recall of 96.5 ± 0.2 percent, a mean average precision at IoU 0.5 ([email protected]) of 98.5 ± 0.3 percent, and a comprehensive [email protected]:0.95 of 84.3 ± 0.2 percent. Compared against the baseline YOLOv11m, these figures represent absolute improvements of 0.4 percentage points in precision, 2.8 percentage points in recall, 1.8 percentage points in [email protected], and 4.6 percentage points in [email protected]:0.95. A paired bootstrap statistical analysis involving 20,000 resamples across matched training seeds produced a 95 percent confidence interval of [0.043, 0.049] for the [email protected]:0.95 gain, demonstrating that the observed performance boost remained strictly positive and consistent across all training runs.
Ablation experiments and leave-one-component-out trials highlighted distinct, specialized roles for each architectural module. In single-module additions to the baseline, C3k2_WTConv drove the largest individual improvement in [email protected]:0.95, lifting the score from 79.7 to 82.0 percent while slightly decreasing overall parameter counts and computational operations. MSCAM produced the strongest individual boost in recall, elevating it to 95.7 percent, while DynamicDCNv4Head yielded the highest [email protected] score of 98.4 percent. Subset evaluations on specific challenge cases further clarified these functional divisions: DySample delivered the highest performance gains on small, distant targets, raising [email protected]:0.95 from 68.1 to 72.8 percent; MSCAM provided the largest gain under foliage occlusion, advancing accuracy from 74.5 to 79.1 percent; DynamicDCNv4Head excelled on curved and twisted fruit geometries, increasing [email protected]:0.95 to 80.7 percent; and both C3k2_WTConv and MSCAM yielded significant gains under green-on-green background conditions.
Comparative evaluations against mainstream and agriculture-specific object detectors positioned FMDD-YOLO at the top of the benchmark in detection and localization accuracy. On the greenhouse dataset, FMDD-YOLO surpassed established models including YOLOv8m (80.4 percent [email protected]:0.95), YOLOv10m (80.9 percent), YOLOv26m (78.7 percent), D-FINE-m (80.6 percent), Faster R-CNN R50-FPN (83.2 percent), and specialized agricultural networks such as DSW-YOLO-m (82.2 percent) and an improved agricultural YOLOv8m (80.5 percent). When retrained on an external public apple-fruitlet dataset containing small, occluded fruits against orchard foliage, FMDD-YOLO demonstrated strong architectural transferability, achieving a leading [email protected]:0.95 of 70.3 ± 0.3 percent. Furthermore, when evaluated directly on the unseen Chili-data validation benchmark without any retraining or parameter fine-tuning, the model maintained an [email protected]:0.95 of 64.2 ± 0.7 percent, outperforming the baseline YOLOv11m by 2.0 percentage points and confirming robust cross-dataset domain generalization.
The study also delineated explicit trade-offs and operational failure boundaries associated with the enhanced architecture. Achieving higher localization accuracy introduced measurable computational overhead: FMDD-YOLO operates with 21.283 million parameters and 72.454 GFLOPs, yielding a forward-pass inference speed of 57.53 frames per second and an end-to-end pipeline latency of 54.041 milliseconds per image (18.504 FPS) on a desktop Nvidia RTX 4090 GPU, compared to 44.873 milliseconds for baseline YOLOv11m. Stage-wise timing confirmed that forward computation within DynamicDCNv4Head constitutes the primary latency bottleneck. The authors emphasize that these desktop evaluations do not establish real-time feasibility on low-power embedded agricultural hardware, noting that future research must explore model pruning, TensorRT acceleration, multi-view camera fusion, and 3D depth sensing to support practical field deployment.
Subject of Research: Computer vision and agricultural object detection for greenhouse green peppers
Article Title: FMDD-YOLO: a frequency-domain multi-scale dynamic and deformable detector for green peppers in complex greenhouse environments
Article References: Zhang, Z., Qin, X., & Li, W. (2026). FMDD-YOLO: a frequency-domain multi-scale dynamic and deformable detector for green peppers in complex greenhouse environments. Journal of King Saud University Computer and Information Sciences, 38(9), Article 1016. https://doi.org/10.1007/s44443-026-01533-5
Image Credits: AI Generated
DOI: 10.1007/s44443-026-01533-5
Keywords: FMDD-YOLO, YOLOv11, green pepper detection, greenhouse agriculture, computer vision, wavelet convolution, WTConv, MSCAM, DySample, DynamicDCNv4Head, deformable convolution, precision agriculture, object detection, deep learning, agricultural automation, horticulture monitoring


1 hour ago
5




















English (US) ·
French (CA) ·