RESUMO
The morphological analysis and volume measurement of the hippocampus are crucial to the study of many brain diseases. Therefore, an accurate hippocampal segmentation method is beneficial for the development of clinical research in brain diseases. U-Net and its variants have become prevalent in hippocampus segmentation of Magnetic Resonance Imaging (MRI) due to their effectiveness, and the architecture based on Transformer has also received some attention. However, some existing methods focus too much on the shape and volume of the hippocampus rather than its spatial information, and the extracted information is independent of each other, ignoring the correlation between local and global features. In addition, many methods cannot be effectively applied to practical medical image segmentation due to many parameters and high computational complexity. To this end, we combined the advantages of CNNs and ViTs (Vision Transformer) and proposed a simple and lightweight model: Light3DHS for the segmentation of the 3D hippocampus. In order to obtain richer local contextual features, the encoder first utilizes a multi-scale convolutional attention module (MCA) to learn the spatial information of the hippocampus. Considering the importance of local features and global semantics for 3D segmentation, we used a lightweight ViT to learn high-level features of scale invariance and further fuse local-to-global representation. To evaluate the effectiveness of encoder feature representation, we designed three decoders of different complexity to generate segmentation maps. Experiments on three common hippocampal datasets demonstrate that the network achieves more accurate hippocampus segmentation with fewer parameters. Light3DHS performs better than other state-of-the-art algorithms.
Assuntos
Hipocampo , Imageamento Tridimensional , Imageamento por Ressonância Magnética , Hipocampo/diagnóstico por imagem , Humanos , Imageamento por Ressonância Magnética/métodos , Imageamento Tridimensional/métodos , Redes Neurais de Computação , Aprendizado Profundo , AlgoritmosRESUMO
Maneuver of conducting polymers (CPs) into lightweight hydrogels can improve their functional performances in energy devices, chemical sensing, pollutant removal, drug delivery, etc. Current approaches for the manipulation of CP hydrogels are limited, and they are mostly accompanied by harsh conditions, tedious processing, compositing with other constituents, or using unusual chemicals. Herein, a two-step route is introduced for the controllable fabrication of CP hydrogels in ambient conditions, where gelation of the shape-anisotropic nano-oxidants followed by in-situ oxidative polymerization leads to the formation of polyaniline (PANI) and polypyrrole hydrogels. The method is readily coupled with different approaches for materials processing of PANI hydrogels into varied shapes, including spherical beads, continuous wires, patterned films, and free-standing objects. In comparison with their bulky counterparts, lightweight PANI items exhibit improved properties when those with specific shapes are used as electrodes for supercapacitors, gas sensors, or dye adsorbents. The current study therefore provides a general and controllable approach for the implementation of CP into hydrogels of varied external shapes, which can pave the way for the integration of lightweight CP structures with emerging functional devices.
RESUMO
There is a growing demand for thermal management materials in electronic fields. Aerogels have attracted interest due to their extremely low density and extraordinary thermal insulation properties. However, the application of aerogels is limited by high production costs and the requirement that aerogel structures not be load-bearing. In this study, mullite-reinforced SiC-based aerogel composite (MR-SiC AC) is prepared through 3D printing combined with in situ growth of SiC nanowires in post processing. The fabricated MR-SiC AC not only has ultra-low thermal conductivity (0.021 W K m-1) and high porosity (90.0%), but also a high Young's modulus (24.4 MPa) and high compressive strength (1.65 MPa), both exceeding the measurements of existing resilient aerogels by an order of magnitude. These properties make MR-SiC AC an ideal solution for the precision thermal management of lightweight structures having complex geometry for functional devices.
RESUMO
Multi-drug combinations for the treatment of complex diseases are gradually becoming an important treatment, and this type of treatment can take advantage of the synergistic effects among drugs. However, drug-drug interactions (DDIs) are not just all beneficial. Accurate and rapid identifications of the DDIs are essential to enhance the effectiveness of combination therapy and avoid unintended side effects. Traditional DDIs prediction methods use only drug sequence information or drug graph information, which ignores information about the position of atoms and edges in the spatial structure. In this paper, we propose Molormer, a method based on a lightweight attention mechanism for DDIs prediction. Molormer takes the two-dimension (2D) structures of drugs as input and encodes the molecular graph with spatial information. Besides, Molormer uses lightweight-based attention mechanism and self-attention distilling to process spatially the encoded molecular graph, which not only retains the multi-headed attention mechanism but also reduces the computational and storage costs. Finally, we use the Siamese network architecture to serve as the architecture of Molormer, which can make full use of the limited data to train the model for better performance and also limit the differences to some extent between networks dealing with drug features. Experiments show that our proposed method outperforms state-of-the-art methods in Accuracy, Precision, Recall and F1 on multi-label DDIs dataset. In the case study section, we used Molormer to make predictions of new interactions for the drugs Aliskiren, Selexipag and Vorapaxar and validated parts of the predictions. Code and models are available at https://github.com/IsXudongZhang/Molormer.
Assuntos
Efeitos Colaterais e Reações Adversas Relacionados a Medicamentos , Interações Medicamentosas , HumanosRESUMO
BACKGROUND: Spine MR image segmentation is important foundation for computer-aided diagnostic (CAD) algorithms of spine disorders. Convolutional neural networks segment effectively, but require high computational costs. PURPOSE: To design a lightweight model based on dynamic level-set loss function for high segmentation performance. STUDY TYPE: Retrospective. POPULATION: Four hundred forty-eight subjects (3163 images) from two separate datasets. Dataset-1: 276 subjects/994 images (53.26% female, mean age 49.02 ± 14.09), all for disc degeneration screening, 188 had disc degeneration, 67 had herniated disc. Dataset-2: public dataset with 172 subjects/2169 images, 142 patients with vertebral degeneration, 163 patients with disc degeneration. FIELD STRENGTH/SEQUENCE: T2 weighted turbo spin echo sequences at 3T. ASSESSMENT: Dynamic Level-set Net (DLS-Net) was compared with four mainstream (including U-net++) and four lightweight models, and manual label made by five radiologists (vertebrae, discs, spinal fluid) used as segmentation evaluation standard. Five-fold cross-validation are used for all experiments. Based on segmentation, a CAD algorithm of lumbar disc was designed for assessing DLS-Net's practicality, and the text annotation (normal, bulging, or herniated) from medical history data were used as evaluation standard. STATISTICAL TESTS: All segmentation models were evaluated with DSC, accuracy, precision, and AUC. The pixel numbers of segmented results were compared with manual label using paired t-tests, with P < 0.05 indicating significance. The CAD algorithm was evaluated with accuracy of lumbar disc diagnosis. RESULTS: With only 1.48% parameters of U-net++, DLS-Net achieved similar accuracy in both datasets (Dataset-1: DSC 0.88 vs. 0.89, AUC 0.94 vs. 0.94; Dataset-2: DSC 0.86 vs. 0.86, AUC 0.93 vs. 0.93). The segmentation results of DLS-Net showed no significant differences with manual labels in pixel numbers for discs (Dataset-1: 1603.30 vs. 1588.77, P = 0.22; Dataset-2: 863.61 vs. 886.4, P = 0.14) and vertebrae (Dataset-1: 3984.28 vs. 3961.94, P = 0.38; Dataset-2: 4806.91 vs. 4732.85, P = 0.21). Based on DLS-Net's segmentation results, the CAD algorithm achieved higher accuracy than using non-cropped MR images (87.47% vs. 61.82%). DATA CONCLUSION: The proposed DLS-Net has fewer parameters but achieves similar accuracy to U-net++, helps CAD algorithm achieve higher accuracy, which facilitates wider application. EVIDENCE LEVEL: 2 TECHNICAL EFFICACY: Stage 1.
Assuntos
Processamento de Imagem Assistida por Computador , Degeneração do Disco Intervertebral , Humanos , Feminino , Adulto , Pessoa de Meia-Idade , Masculino , Processamento de Imagem Assistida por Computador/métodos , Estudos Retrospectivos , Degeneração do Disco Intervertebral/diagnóstico por imagem , Redes Neurais de Computação , Coluna Vertebral/diagnóstico por imagemRESUMO
The fracture behaviour of concrete is studied in various micro- and macro-damage models. This is important for estimating serviceability and stability of concrete structures. However, a detailed understanding of the material behaviour under load is often not available. In order to better interpret the fracture behaviour and pattern, images of lightweight concrete were taken using a high-resolution computed tomography (µ-CT) scanner. The samples were loaded between the taken images and the load was kept constant during the measurement. This study describes the method used and how the data set was analysed to investigate displacements and cracks. It has been shown that displacements and damage to the concrete structure can be detected prior to failure, allowing conclusions to be drawn about the structural behaviour. In principle, the µ-CT measurement can be used to examine different kinds of concrete as well as other systems with inorganic binders and to compare the fracture behaviour of different systems.
RESUMO
OBJECTIVE: To evaluate the performance of two lightweight neural network models in the diagnosis of common fundus diseases and make comparison to another two classical models. METHODS: A total of 16,000 color fundus photography were collected, including 2000 each of glaucoma, diabetic retinopathy (DR), high myopia, central retinal vein occlusion (CRVO), age-related macular degeneration (AMD), optic neuropathy, and central serous chorioretinopathy (CSC), in addition to 2000 normal fundus. Fundus photography was obtained from patients or physical examiners who visited the Ophthalmology Department of Beijing Tongren Hospital, Capital Medical University. Each fundus photography has been diagnosed and labeled by two professional ophthalmologists. Two classical classification models (ResNet152 and DenseNet121), and two lightweight classification models (MobileNetV3 and ShufflenetV2), were trained. Area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value, and negative predictive value were used to evaluate the performance of the four models. RESULTS: Compared with the classical classification model, the total size and number of parameters of the two lightweight classification models were significantly reduced, and the classification speed was sharply improved. Compared with the DenseNet121 model, the ShufflenetV2 model took 50.7% less time to make a diagnosis on a fundus photography. The classical models performed better than lightweight classification models, and Densenet121 showed highest AUC in five out of the seven common fundus diseases. However, the performance of lightweight classification models is satisfying. The AUCs using MobileNetV3 model to diagnose AMD, diabetic retinopathy, glaucoma, CRVO, high myopia, optic atrophy, and CSC were 0.805, 0.892, 0.866, 0.812, 0.887, 0.868, and 0.803, respectively. For ShufflenetV2model, the AUCs for the above seven diseases were 0.856, 0.893, 0.855, 0.884, 0.891, 0.867, and 0.844, respectively. CONCLUSION: The training of light-weight neural network models based on color fundus photography for the diagnosis of common fundus diseases is not only fast but also has a significant reduction in storage size and parameter number compared with the classical classification model, and can achieve satisfactory accuracy.
Assuntos
Retinopatia Diabética , Glaucoma , Degeneração Macular , Miopia , Humanos , Retinopatia Diabética/diagnóstico , Técnicas de Diagnóstico Oftalmológico , Fundo de Olho , Glaucoma/diagnóstico , Degeneração Macular/diagnóstico , FotografaçãoRESUMO
BACKGROUND: Due to the declining mortality rates of breast carcinoma and the rising incidence of risk-reducing mastectomies, enhancing the quality of life after breast reconstructions has become an increasingly important goal. The advantages of lightweight breast implants (B-Lite®) may significantly contribute to achieving this objective. This study aims to investigate whether lightweight implants are suitable for patients undergoing breast reconstruction and could improve the quality of life in comparison to conventional implants. METHODS: In this study, we retrospectively analyzed 48 patients (38 implants in each group) who underwent implant-based breast reconstruction with either B-Lite® or conventional breast implants between 2019 and 2022 at the University Center for Plastic Surgery in Regensburg. As part of the postoperative follow-up, a clinical examination and a survey using the Breast-Q® questionnaire were conducted to evaluate the postoperative quality of life. RESULTS: The implants used were similar in weight and shape. On average, the B-Lite® implants had a higher implant volume and patients in this group had a slightly higher BMI. Patients who received B-Lite® implants showed a significantly better result regarding the sensation of sensitivity in the surgical area and the scar formation also appeared to be more favorable. However, patients with B-Lite® implants perceived their implants as more uncomfortable than those with conventional breast implants. In other terms concerning quality of life, both groups appeared similar. CONCLUSION: In summary, there are confounding factors that could influence the outcome of some aspects in this study, which could not be avoided due to the retrospective study design and the temporary suspension of B-Lite implants. Nevertheless, as the first of its kind, this study demonstrated that B-Lite implants could also be suitable for usage in breast reconstructions, thus providing an important foundation for further prospective studies to build upon.
RESUMO
To address the lightweight and real-time issues of coal sorting detection, an intelligent detection method for coal and gangue, Our-v8, was proposed based on improved YOLOv8. Images of coal and gangue with different densities under two diverse lighting environments were collected. Then the Laplacian image enhancement algorithm was proposed to improve the training data quality, sharpening contours and boosting feature extraction; the CBAM attention mechanism was introduced to prioritize crucial features, enhancing more accurate feature extraction ability; and the EIOU loss function was added to refine box regression, further improving detection accuracy. The experimental results showed that Our-v8 for detecting coal and gangue in a halogen lamp lighting environment achieved excellent performance with a mean average precision (mAP) of 99.5%, was lightweight with FLOPs of 29.7, Param of 12.8, and a size of only 22.1 MB. Additionally, Our-v8 can provide accurate location information for coal and gangue, making it ideal for real-time coal sorting applications.
RESUMO
Considering the high incidence of accidents at tunnel construction sites, using robots to replace humans in hazardous tasks can effectively safeguard their lives. However, most robots currently used in this field require manual control and lack autonomous obstacle avoidance capability. To address these issues, we propose a lightweight model based on an improved version of YOLOv5 for obstacle detection. Firstly, to enhance detection speed and reduce computational load, we modify the backbone network to the lightweight Shufflenet v2. Secondly, we introduce a coordinate attention mechanism to enhance the network's ability to learn feature representations. Subsequently, we replace the neck convolution block with GSConv to improve the model's efficiency. Finally, we modify the model's upsampling method to further enhance detection accuracy. Through comparative experiments on the model, the results demonstrate that our approach achieves an approximately 37% increase in detection speed with a minimal accuracy reduction of 1.5%. The frame rate has improved by about 54%, the parameter count has decreased by approximately 74%, and the model size has decreased by 2.5 MB. The experimental results indicate that our method can reduce hardware requirements for the model, striking a balance between detection speed and accuracy.
RESUMO
With the continuous development of deep learning, the application of object detection based on deep neural networks in the coal mine has been expanding. Simultaneously, as the production applications demand higher recognition accuracy, most research chooses to enlarge the depth and parameters of the network to improve accuracy. However, due to the limited computing resources in the coal mining face, it is challenging to meet the computation demands of a large number of hardware resources. Therefore, this paper proposes a lightweight object detection algorithm designed specifically for the coal mining face, referred to as CM-YOLOv8. The algorithm introduces adaptive predefined anchor boxes tailored to the coal mining face dataset to enhance the detection performance of various targets. Simultaneously, a pruning method based on the L1 norm is designed, significantly compressing the model's computation and parameter volume without compromising accuracy. The proposed algorithm is validated on the coal mining dataset DsLMF+, achieving a compression rate of 40% on the model volume with less than a 1% drop in accuracy. Comparative analysis with other existing algorithms demonstrates its efficiency and practicality in coal mining scenarios. The experiments confirm that CM-YOLOv8 significantly reduces the model's computational requirements and volume while maintaining high accuracy.
RESUMO
A vehicle detection algorithm is an indispensable component of intelligent traffic management and control systems, influencing the efficiency and functionality of the system. In this paper, we propose a lightweight improvement method for the YOLOv5 algorithm based on integrated perceptual attention, with few parameters and high detection accuracy. First, we propose a lightweight module IPA with a Transformer encoder based on integrated perceptual attention, which leads to a reduction in the number of parameters while capturing global dependencies for richer contextual information. Second, we propose a lightweight and efficient multiscale spatial channel reconstruction (MSCCR) module that does not increase parameter and computational complexity and facilitates representative feature learning. Finally, we incorporate the IPA module and the MSCCR module into the YOLOv5s backbone network to reduce model parameters and improve accuracy. The test results show that, compared with the original model, the model parameters decrease by about 9%, the average accuracy (mAP@50) increases by 3.1%, and the FLOPS does not increase.
RESUMO
Vehicle detection is a research direction in the field of target detection and is widely used in intelligent transportation, automatic driving, urban planning, and other fields. To balance the high-speed advantage of lightweight networks and the high-precision advantage of multiscale networks, a vehicle detection algorithm based on a lightweight backbone network and a multiscale neck network is proposed. The mobile NetV3 lightweight network based on deep separable convolution is used as the backbone network to improve the speed of vehicle detection. The icbam attention mechanism module is used to strengthen the processing of the vehicle feature information detected by the backbone network to enrich the input information of the neck network. The bifpn and icbam attention mechanism modules are integrated into the neck network to improve the detection accuracy of vehicles of different sizes and categories. A vehicle detection experiment on the Ua-Detrac dataset verifies that the proposed algorithm can effectively balance vehicle detection accuracy and speed. The detection accuracy is 71.19%, the number of parameters is 3.8 MB, and the detection speed is 120.02 fps, which meets the actual requirements of the parameter quantity, detection speed, and accuracy of the vehicle detection algorithm embedded in the mobile device.
RESUMO
X-ray images typically contain complex background information and abundant small objects, posing significant challenges for object detection in security tasks. Most existing object detection methods rely on complex networks and high computational costs, which poses a challenge to implement lightweight models. This article proposes Fine-YOLO to achieve rapid and accurate detection in the security domain. First, a low-parameter feature aggregation (LPFA) structure is designed for the backbone feature network of YOLOv7 to enhance its ability to learn more information with a lighter structure. Second, a high-density feature aggregation (HDFA) structure is proposed to solve the problem of loss of local details and deep location information caused by the necked feature fusion network in YOLOv7-Tiny-SiLU, connecting cross-level features through max-pooling. Third, the Normalized Wasserstein Distance (NWD) method is employed to alleviate the convergence complexity resulting from the extreme sensitivity of bounding box regression to small objects. The proposed Fine-YOLO model is evaluated on the EDS dataset, achieving a detection accuracy of 58.3% with only 16.1 M parameters. In addition, an auxiliary validation is performed on the NEU-DET dataset, the detection accuracy reaches 73.1%. Experimental results show that Fine-YOLO is not only suitable for security, but can also be extended to other inspection areas.
RESUMO
As an important direction in computer vision, human pose estimation has received extensive attention in recent years. A High-Resolution Network (HRNet) can achieve effective estimation results as a classical human pose estimation method. However, the complex structure of the model is not conducive to deployment under limited computer resources. Therefore, an improved Efficient and Lightweight HRNet (EL-HRNet) model is proposed. In detail, point-wise and grouped convolutions were used to construct a lightweight residual module, replacing the original 3 × 3 module to reduce the parameters. To compensate for the information loss caused by the network's lightweight nature, the Convolutional Block Attention Module (CBAM) is introduced after the new lightweight residual module to construct the Lightweight Attention Basicblock (LA-Basicblock) module to achieve high-precision human pose estimation. To verify the effectiveness of the proposed EL-HRNet, experiments were carried out using the COCO2017 and MPII datasets. The experimental results show that the EL-HRNet model requires only 5 million parameters and 2.0 GFlops calculations and achieves an AP score of 67.1% on the COCO2017 validation set. In addition, PCKh@0.5mean is 87.7% on the MPII validation set, and EL-HRNet shows a good balance between model complexity and human pose estimation accuracy.
RESUMO
Convolutional neural networks (CNNs) have made significant progress in the field of facial expression recognition (FER). However, due to challenges such as occlusion, lighting variations, and changes in head pose, facial expression recognition in real-world environments remains highly challenging. At the same time, methods solely based on CNN heavily rely on local spatial features, lack global information, and struggle to balance the relationship between computational complexity and recognition accuracy. Consequently, the CNN-based models still fall short in their ability to address FER adequately. To address these issues, we propose a lightweight facial expression recognition method based on a hybrid vision transformer. This method captures multi-scale facial features through an improved attention module, achieving richer feature integration, enhancing the network's perception of key facial expression regions, and improving feature extraction capabilities. Additionally, to further enhance the model's performance, we have designed the patch dropping (PD) module. This module aims to emulate the attention allocation mechanism of the human visual system for local features, guiding the network to focus on the most discriminative features, reducing the influence of irrelevant features, and intuitively lowering computational costs. Extensive experiments demonstrate that our approach significantly outperforms other methods, achieving an accuracy of 86.51% on RAF-DB and nearly 70% on FER2013, with a model size of only 3.64 MB. These results demonstrate that our method provides a new perspective for the field of facial expression recognition.
Assuntos
Expressão Facial , Redes Neurais de Computação , Humanos , Reconhecimento Facial Automatizado/métodos , Algoritmos , Processamento de Imagem Assistida por Computador/métodos , Face , Reconhecimento Automatizado de Padrão/métodosRESUMO
In the field of autofocus for optical systems, although passive focusing methods are widely used due to their cost-effectiveness, fixed focusing windows and evaluation functions in certain scenarios can still lead to focusing failures. Additionally, the lack of datasets limits the extensive research of deep learning methods. In this work, we propose a neural network autofocus method with the capability of dynamically selecting the region of interest (ROI). Our main work is as follows: first, we construct a dataset for automatic focusing of grayscale images; second, we transform the autofocus issue into an ordinal regression problem and propose two focusing strategies: full-stack search and single-frame prediction; and third, we construct a MobileViT network with a linear self-attention mechanism to achieve automatic focusing on dynamic regions of interest. The effectiveness of the proposed focusing method is verified through experiments, and the results show that the focusing MAE of the full-stack search can be as low as 0.094, with a focusing time of 27.8 ms, and the focusing MAE of the single-frame prediction can be as low as 0.142, with a focusing time of 27.5 ms.
RESUMO
The identification of safflower filament targets and the precise localization of picking points are fundamental prerequisites for achieving automated filament retrieval. In light of challenges such as severe occlusion of targets, low recognition accuracy, and the considerable size of models in unstructured environments, this paper introduces a novel lightweight YOLO-SaFi model. The architectural design of this model features a Backbone layer incorporating the StarNet network; a Neck layer introducing a novel ELC convolution module to refine the C2f module; and a Head layer implementing a new lightweight shared convolution detection head, Detect_EL. Furthermore, the loss function is enhanced by upgrading CIoU to PIoUv2. These enhancements significantly augment the model's capability to perceive spatial information and facilitate multi-feature fusion, consequently enhancing detection performance and rendering the model more lightweight. Performance evaluations conducted via comparative experiments with the baseline model reveal that YOLO-SaFi achieved a reduction of parameters, computational load, and weight files by 50.0%, 40.7%, and 48.2%, respectively, compared to the YOLOv8 baseline model. Moreover, YOLO-SaFi demonstrated improvements in recall, mean average precision, and detection speed by 1.9%, 0.3%, and 88.4 frames per second, respectively. Finally, the deployment of the YOLO-SaFi model on the Jetson Orin Nano device corroborates the superior performance of the enhanced model, thereby establishing a robust visual detection framework for the advancement of intelligent safflower filament retrieval robots in unstructured environments.
RESUMO
Aiming at the problems of target detection models in traffic scenarios including a large number of parameters, heavy computational burden, and high application cost, this paper introduces an enhanced lightweight real-time detection algorithm, which exhibits higher detection speed and accuracy for vehicle detection. This paper considers the YOLOv7 algorithm as the benchmark model, designs a lightweight backbone network, and uses the MobileNetV3 lightweight network to extract target features. Inspired by the structure of SPPF, the spatial pyramid pooling module is reconfigured by incorporating GSConv, and a lightweight SPPFCSPC-GS module is designed, aiming to minimize the quantity of model parameters and enhance the training speed even further. Furthermore, the CA mechanism is integrated to enhance the feature extraction capability of the model. Finally, the MPDIoU loss function is utilized to optimize the model's training process. Experiments showcase that the refined YOLOv7 algorithm can achieve 98.2% mAP on the BIT-Vehicle dataset with 52.8% fewer model parameters than the original model and a 35.2% improvement in FPS. The enhanced model adeptly strikes a finer equilibrium between velocity and precision, providing favorable conditions for embedding the model into mobile devices.
RESUMO
Wheat seed detection has important applications in calculating thousand-grain weight and crop breeding. In order to solve the problems of seed accumulation, adhesion, and occlusion that can lead to low counting accuracy, while ensuring fast detection speed with high accuracy, a wheat seed counting method is proposed to provide technical support for the development of the embedded platform of the seed counter. This study proposes a lightweight real-time wheat seed detection model, YOLOv8-HD, based on YOLOv8. Firstly, we introduce the concept of shared convolutional layers to improve the YOLOv8 detection head, reducing the number of parameters and achieving a lightweight design to improve runtime speed. Secondly, we incorporate the Vision Transformer with a Deformable Attention mechanism into the C2f module of the backbone network to enhance the network's feature extraction capability and improve detection accuracy. The results show that in the stacked scenes with impurities (severe seed adhesion), the YOLOv8-HD model achieves an average detection accuracy (mAP) of 77.6%, which is 9.1% higher than YOLOv8. In all scenes, the YOLOv8-HD model achieves an average detection accuracy (mAP) of 99.3%, which is 16.8% higher than YOLOv8. The memory size of the YOLOv8-HD model is 6.35 MB, approximately 4/5 of YOLOv8. The GFLOPs of YOLOv8-HD decrease by 16%. The inference time of YOLOv8-HD is 2.86 ms (on GPU), which is lower than YOLOv8. Finally, we conducted numerous experiments and the results showed that YOLOv8-HD outperforms other mainstream networks in terms of mAP, speed, and model size. Therefore, our YOLOv8-HD can efficiently detect wheat seeds in various scenarios, providing technical support for the development of seed counting instruments.