Neural Network Classification Model Research Articles

Research of 2D-COS with metabolomics modifications through deep learning for traceability of wine

To tackle the difficulty of extracting features from one-dimensional spectral signals using traditional spectral analysis, a metabolomics analysis method is proposed to locate two-dimensional correlated spectral feature bands and combine it with deep learning classification for wine origin traceability. Metabolomics analysis was performed on 180 wine samples from 6 different wine regions using UPLC-Q-TOF-MS. Indole, Sulfacetamide, and caffeine were selected as the main differential components. By analyzing the molecular structure of these components and referring to the main functional groups on the infrared spectrum, characteristic band regions with wavelengths in the range of 1000–1400 nm and 1500–1800 nm were selected. Draw two-dimensional correlation spectra (2D-COS) separately, generate synchronous correlation spectra and asynchronous correlation spectra, establish convolutional neural network (CNN) classification models, and achieve the purpose of wine origin traceability. The experimental results demonstrate that combining two segments of two-dimensional characteristic spectra determined by metabolomics screening with convolutional neural networks yields optimal classification results. This validates the effectiveness of using metabolomics screening to determine spectral feature regions in tracing wine origin. This approach effectively removes irrelevant variables while retaining crucial chemical information, enhancing spectral resolution. This integrated approach strengthens the classification model's understanding of samples, significantly increasing accuracy.

Accelerating the Screening of Modified MA2Z4 Catalysts for Hydrogen Evolution Reaction by Deep Learning‐Based Local Geometric Analysis

Machine learning (ML) integrated with density functional theory (DFT) calculations have recently been used to accelerate the design and discovery of single‐atom catalysts (SACs) by establishing deep structure–activity relationships. The traditional ML models are always difficult to identify the structural differences among the single‐atom systems with different modification methods, leading to the limitation of the potential application range. Aiming to the structural properties of several typical two‐dimensional MA2Z4‐based single‐atom systems (bare MA2Z4 and metal single‐atom doped/supported MA2Z4), an improved crystal graph convolutional neural network (CGCNN) classification model was employed, instead of the traditional machine learning regression model, to address the challenge of incompatibility in the studied systems. The CGCNN model was optimized using crystal graph representation in which the geometric configuration was divided into active layer, surface layer, and bulk layer (ASB‐GCNN). Through ML and DFT calculations, five potential single‐atom hydrogen evolution reaction (HER) catalysts were screened from chemical space of 600 MA2Z4‐based materials, especially V1/HfSn2N4(S) with high stability and activity (ΔGH* is 0.06 eV). Further projected density of states (pDOS) analysis in combination with the wave function analysis of the SAC‐H bond revealed that the SAC‐dz2 orbital coincided with the H‐s orbital around the energy level of −2.50 eV, and orbital analysis confirmed the formation of σ bonds. This study provides an efficient multistep screening design framework of metal single‐atom catalyst for HER systems with similar two‐dimensional supports but different geometric configurations.

Improved Artificial Neural Network Classification Model based Metaheuristic Optimization for Handwritten Character Recognition

This study addresses the concerns regarding the performance of Handwritten Character Recognition (HCR) systems, focusing on the classification stage. It is widely acknowledged that the development of the classification model significantly impacts the overall performance of HCR. The problems identified specifically pertain to the classification model, particularly in the context of the Artificial Neural Network (ANN) learning problem, leading to low accuracy in recognizing handwritten characters. The objective of this study is to improve and refine the ANN classification model to achieve better HCR. To achieve this goal, this study proposed a hybrid Flower Pollination Algorithm with Artificial Neural Network (FPA-ANN) classification model for HCR. The FPA is one of the metaheuristic approaches is utilized as an optimization technique to enhance the performance of ANN, particularly by optimizing the network training process of ANN. The experimentation phase involves using the National Institute of Standards and Technology (NIST) handwritten character database. Finally, the proposed FPA-ANN classification model is analysed based on generated confusion matrix and evaluated performance of the classification model in terms of precision, sensitivity, specificity, F-score and accuracy.

Visualisation of moisture content distribution maps and classification of freshness level of loquats

Moisture content (MC) and freshness undergo evident changes during the storage of loquat fruits and thus are important indices of loquat fruit quality. Conventional methods for assessing MC and freshness are typically destructive and time-consuming, and cannot provide distribution maps of MC in the whole loquats. In this study, the use of hyperspectral imaging (HSI) was investigated for rapid and non-destructive prediction of MC in loquats, and for the visualisation of their MC distribution maps. Furthermore, freshness level discrimination models for loquats at different storage times were established using partial least squares discriminant analysis (PLS-DA), multinomial logistic regression (MLogR), and back-propagation neural network (BPNN) classification models. MC was well-predicted, with the best residual predictive deviation (RPD) of 2.40. Using the best prediction model, the distribution maps of MC in loquats were visualised. Additionally, the competitive adaptive reweighted sampling (CARS)-BPNN model obtained superior discrimination performance, with 95.56 % and 93.33 % accuracy for the calibration and prediction sets. Our results show that HSI is a promising fast and non-destructive method for determining MC and freshness of loquats.

Detection and metrology of micro-nano particles through the integration of image and laser modes

Micro-nano particles are an indispensable component in various industrial and scientific applications. Accurate granular information is critical for manufacturing products involving these particles. This study presents a novel detection method that merges image and laser modes to achieve precise size and shape determination of micro-nano particles. The proposed method integrates adaptive image analysis with laser particle size fitting to ensure accurate feature detection. Experimental results using ternary precursor samples demonstrate that the fusion mode outperforms both the laser and image modes in terms of size and shape accuracy. The errors of D10, D50, and D90 of the fusion mode are 4.89%, 1.88%, and 2.94%, respectively, which are significantly lower than those obtained using the laser or image mode alone. Additionally, the fusion mode retains the particle shape information from the image mode. The fusion model offers a promising approach for dual-mode detection of micro-nano particles, demonstrating enhanced robustness compared to single-mode methods. Furthermore, the study introduces an image processing method based on a BP neural network classification model that accurately classifies and detects agglomerated and non-agglomerated particles. The classification accuracy reaches 94.98%, and the particle size information remains within the acceptable range.

Diabetic retinopathy detection using Bilayered Neural Network classification model with resubstitution validation

In recent years, eye diseases in diabetic patients are one of the most common has been diabetic retinopathy (DR). which leads to complete blindness in advanced stages. Diabetes affects the blood vessels in the retina and causes vision loss. One of the ways to decrease the risk of this issue is to detect diabetic retinopathy in its early stages. This study describes a computer-aided screening system (DREAM) that uses a neural network classification model in machine learning to assess fundus images with different illumination and fields of vision and provide a severity grade for diabetic retinopathy. Moreover, the methodology of this study based on:•Enhancement techniques have been used on dataset images, histogram equalization, noise reduction and image scaling,•vSLAM has been selected as feature extraction,•Bilayered Neural Network under resubstitution validation used as a classification model.Finally, after testing on the DR severity grading system is tested on 6332 images of detection of diabetic retinopathy images, the result of the ROC curve is 0.985 on image dataset and obtains accuracy reached 98.5%. The classification result has been founded under MATLAB platform, beside it that be work with real time analysis and detection when patient eyes analysis.

A resource-efficient quantum convolutional neural network

Quantum Convolutional Neural Network (QCNN) has achieved significant success in solving various complex problems, such as quantum many-body physics and image recognition. In comparison to the classical Convolutional Neural Network (CNN) model, the QCNN model requires excellent numerical performance or efficient computational resources to showcase its potential quantum advantages, particularly in classical data processing tasks. In this paper, we propose a computationally resource-efficient QCNN model referred to as RE-QCNN. Specifically, through a comprehensive analysis of the complexity associated with the forward and backward propagation processes in the quantum convolutional layer, our results demonstrate a significant reduction in computational resources required for this layer compared to the classical CNN model. Furthermore, our model is numerically benchmarked on recognizing images from the MNIST and Fashion-MNIST datasets, achieving high accuracy in these multi-class classification tasks.

Early detection of infestation by mustard aphid, vegetable thrips and two-spotted spider mite in bok choy with deep neural network (DNN) classification model using hyperspectral imaging data

Although exclusion measures (e.g., air filters, biosecurity practices) can be employed to prevent occurrence of pest outbreaks, indoors vegetable farms in Singapore are still susceptible to various arthropod pests. Due to strong interest from the industry to pursue pesticide-free production, indoors pest management is often focused on early detection for timely containment and eradication, implying the importance of robust and vigorous pest monitoring programs. In recent years, application of machine vision technologies, especially hyperspectral imaging (HSI), has been studied for their capacity to early detect pest infestation. However, there is a lack of studies conducted in actual indoor environments and on multiple arthropod pests. Thus, this study aimed to non-destructively collect hyperspectral data of bok choy Brassica rapa subspecies chinensis which were healthy or infested with either mustard aphids Lipaphis erisymi, vegetable thrips Echinothrips americanus or two-spotted spider mites Tetranichus urticae in indoor environment to build deep neural network (DNN) classification model for early detection. Based on HSI data of control and infested plants collected daily over a period of two weeks, we found that point percentage change (PPC) values associated with leaf reflectance in 420–440 nm, 500–520 nm, 620–637 nm, 720–800 nm, and 850 nm were sensitive to infestation by the mentioned arthropod pests. Deep Neural Network (DNN) classification models trained on collected HSI data were found to outperform Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM) classification models. DNN models achieved 92.8 ± 0.4 % overall classification accuracy across all days. As early as two days after infestation, DNN models could achieved classification precision values of 96.4 %, 96.9 %, 93.9 % and 100 % for control plants and plants infested with either aphids, spider mites or thrips respectively. These results highlight the feasibility of multiclass early detection of different arthropod pests and the potential of HSI system coupled with DNN classification as an autonomous plant health monitoring tool in indoor crop production.

High-throughput thermoelectric materials screening by deep convolutional neural network with fused orbital field matrix and composition descriptors

Thermoelectric materials harvest waste heat and convert it into reusable electricity. Thermoelectrics are also widely used in inverse ways such as refrigerators and cooling electronics. However, most popular and known thermoelectric materials to date were proposed and found by intuition, mostly through experiments. Unfortunately, it is extremely time and resource consuming to synthesize and measure the thermoelectric properties through trial-and-error experiments. Here, we develop a convolutional neural network (CNN) classification model that utilizes the fused orbital field matrix and composition descriptors to screen a large pool of materials to discover new thermoelectric candidates with power factor higher than 10 μW/cm K2. The model used our own data generated by high-throughput density functional theory calculations coupled with ab initio scattering and transport package to obtain electronic transport properties without assuming constant relaxation time of electrons, which ensures more reliable electronic transport properties calculations than previous studies. The classification model was also compared to some traditional machine learning algorithms such as gradient boosting and random forest. We deployed the classification model on 3465 cubic dynamically stable structures with non-zero bandgap screened from Open Quantum Materials Database. We identified many high-performance thermoelectric materials with ZT > 1 or close to 1 across a wide temperature range from 300 to 700 K and for both n- and p-type doping with different doping concentrations. Moreover, our feature importance and maximal information coefficient analysis demonstrates two previously unreported material descriptors, namely, mean melting temperature and low average deviation of electronegativity, that are strongly correlated with power factor and thus provide a new route for quickly screening potential thermoelectrics with high success rate. Our deep CNN model with fused orbital field matrix and composition descriptors is very promising for screening high power factor thermoelectrics from large-scale hypothetical structures.

Multi-fault diagnosis of district heating system based on PCA_BP neural network

To diagnose multiple faults (blockage and leakage) of a district heating system (MFDHS) in a timely and accurate manner, a new multi-fault diagnosis method is proposed that considers fault diagnosis as a classification problem based on principal component analysis (PCA) and a Back Propagation (BP) neural network. Using this method, a simulation model of heating-network faults is constructed and fault experiments are conducted. Based on the simulation model and experiments, the experimental (ED) and model (MD) datasets are effectively constructed followed by the construction of substitute datasets(SD) that replaces the experimental data with model data. PCA and data standardization are then adopted to process all the datasets. Finally, three BP neural network classification models are trained and tested using the three datasets. To demonstrate the effectiveness of this method, four experimental heating networks, namely a single heat source and branch (SHB), single heat source and single loop (SHSL), double heat source and branch (DHB), and double heat source and double loop (DHDL), are discussed as case studies. The results show that the relative errors of the experimental and simulation data for the pressures and flow rates are less than 5%. The highest prediction accuracy is obtained when the fault to the normal data are 95–5% under training sets to testing sets of 70–30%, while for the four heating networks, the prediction accuracy of the method built with ED is 99.93%, 99.99%, 99.84%, and 99.66% respectively, whereas it reaches 99.99% for MD. The accuracies are 99.99%, 98.13%, 97.4%, and 99.99%, respectively, with MD serving as the training set while another 11400 sets of experimental data serve as the testing set. Accordingly, the accuracy ranges from 95.56% to 88.03%, 97.85–93.29%, 95.65–85.71%, and 100–84.82%, with SD serving as the training set and another 11400 sets of experimental data serving as the testing set.

Multi-sensor fusion based optimized deep convolutional neural network for boxing punch activity recognition

Recent advancements in deep learning have significantly enhanced the recognition of player activities in sports by enabling automatic feature extraction. In our proposed work, we focus on recognizing six distinct punches in the context of boxing. We incorporate the sliding window technique during the pre-processing stage to transform the time-series data obtained from Inertial Measurement Unit (IMU) sensors in a boxing punch activity recognition system. Our approach influences a sensor fusion-based Deep Convolutional Neural Network (DCNN) classification model to identify various boxing punches accurately, achieving an impressive F1 score. The system demonstrates proficiency in distinguishing similar activities, such as jab and hook punches where the existing systems made misclassifications due to subtle variations in arm flexion that differentiate the two. Through experimentation, we identify an optimal window size for boxing punch activity recognition, which falls within the range of 15–20 frames (equivalent to 0.15–0.25 s). This window size selection results in a notable reduction in inference time. To evaluate our proposed model, we conduct comparisons with a standard DCNN and an optimized DCNN model. Our proposed optimized DCNN model demonstrates enhanced recognition accuracy, achieving an impressive 99%, coupled with an improved F1 score of 87%. Furthermore, the model displays a remarkable reduction in inference time, clocking in at less than 1 ms. Overall, our research contributes to the field of sports-related player activity recognition by employing the power of deep learning. By expertly combining these techniques, we achieve remarkable accuracy, precision, and efficiency in recognizing various boxing punches.

Performance of convolutional neural networks on CIFAR and Dog Breed classification dataset

This research aims to study the deep learning applications in image recognition and classification tasks. The advantages and limitations of deep learning are explored by analysing existing deep learning algorithms and their applications on tasks such as using images to do classification, image segmentation, and target detection. In the experimental part, this paper evaluates the performance of deep learning in image recognition by using classical neural network models, including training and testing models on a large size of image datasets. The results show that neural network has good classification and detection capabilities in image recognition tasks and also achieves good results for image segmentation tasks. However, the model training process of shallow neural network models is time-consuming and performs poorly for small-scale datasets. Based on these situations, this paper proposes some optimization strategies to achieve high performance and efficiency of neural networks in image recognition and analyze the performance differences between different strategies .

Abnormal data detection and recovery of sensors network based on spatiotemporal deep learning methodology

This paper proposes a novel deep-learning-enabled framework for abnormal data detection and recovery considering spatial-temporal correlation among sensors. The wavelet scattering transform is adopted to replace the first layer of one-dimension convolutional neural network classification model. The multiple typical sensor failure modes of sensor are identified including bias, drifting, gain, precision degradation, and three complete faults (constant; constant + noise; noise). The bidirectional long-short time memory (BLSTM) network is developed to recover abnormal data through positive and negative propagation. The BLSTM cell involving forget, input, and output gate is utilized to adaptively transmit valuable information of long-term data. Different sensor configurations are conducted to select the optimal input sensors of BLSTM model. The proposed models are then validated by three engineering cases. The results show that the failure mode diagnosis and recovery accuracy are over 90 %. Sensor selection has a large impact on the data reconstruction, and selecting sensors with higher correlation can achieve a higher accuracy. In addition, the data recovery accuracy of the proposed BLSTM model surpasses that of traditional long-short time memory models.

Deep hybrid modeling of a HEK293 process: Combining long short-term memory networks with first principles equations.

The combination of physical equations with deep learning is becoming a promising methodology for bioprocess digitalization. In this paper, we investigate for the first time the combination of long short-term memory (LSTM) networks with first principles equations in a hybrid workflow to describe human embryonic kidney 293 (HEK293) culture dynamics. Experimental data of 27 extracellular state variables in 20 fed-batch HEK293 cultures were collected in a parallel high throughput 250 mL cultivation system in an industrial process development setting. The adaptive moment estimation method with stochastic regularization and cross-validation were employed for deep learning. A total of 784 hybrid models with varying deep neural network architectures, depths, layers sizes and node activation functions were compared. In most scenarios, hybrid LSTM models outperformed classical hybrid Feedforward Neural Network (FFNN) models in terms of training and testing error. Hybrid LSTM models revealed to be less sensitive to data resampling than FFNN hybrid models. As disadvantages, Hybrid LSTM models are in general more complex (higher number of parameters) and have a higher computation cost than FFNN hybrid models. The hybrid model with the highest prediction accuracy consisted in a LSTM network with seven internal states connected in series with dynamic material balance equations. This hybrid model correctly predicted the dynamics of the 27 state variables (R2 = 0.93 in the test data set), including biomass, key substrates, amino acids and metabolic by-products for around 10 cultivation days.

PathoGraph: An Attention-Based Graph Neural Network Capable of Prognostication Based on CD276 Labelling of Malignant Glioma Cells.

Computerized methods have been developed that allow quantitative morphological analyses of whole slide images (WSIs), e.g., of immunohistochemical stains. The latter are attractive because they can provide high-resolution data on the distribution of proteins in tissue. However, many immunohistochemical results are complex because the protein of interest occurs in multiple locations (in different cells and also extracellularly). We have recently established an artificial intelligence framework, PathoFusion which utilises a bifocal convolutional neural network (BCNN) model for detecting and counting arbitrarily definable morphological structures. We have now complemented this model by adding an attention-based graph neural network (abGCN) for the advanced analysis and automated interpretation of such data. Classical convolutional neural network (CNN) models suffer from limitations when handling global information. In contrast, our abGCN is capable of creating a graph representation of cellular detail from entire WSIs. This abGCN method combines attention learning with visualisation techniques that pinpoint the location of informative cells and highlight cell-cell interactions. We have analysed cellular labelling for CD276, a protein of great interest in cancer immunology and a potential marker of malignant glioma cells/putative glioma stem cells (GSCs). We are especially interested in the relationship between CD276 expression and prognosis. The graphs permit predicting individual patient survival on the basis of GSC community features. Our experiments lay a foundation for the use of the BCNN-abGCN tool chain in automated diagnostic prognostication using immunohistochemically labelled histological slides, but the method is essentially generic and potentially a widely usable tool in medical research and AI based healthcare applications.

Status Identification in Support of Fishing Effort Estimation for Tuna Longliners in Waters near the Marshall Islands Based on AIS Data

Visualising the fishing behaviour of vessels and quantifying the spatial distribution of fishing effort is the scientific basis for assessing and managing fisheries resources. The information on the dynamics of fishing vessel voyages provided by the automatic identification system (AIS) of vessels serves as high-precision fishery data and provides a means of quantifying fishing effort with high spatial and temporal resolution in the tuna longline fishery. Based on the AIS data of five tuna longliners operating in the waters near the Marshall Islands from 2020 to 2021, this study used three methods, namely the threshold screening method, the construction of a BP neural network and the support vector machine (SVM) to identify the fishing and non-fishing status of the tuna longliners, respectively. This study investigates the status identification and fishing effort estimation of the tuna longliner (VESSEL A) in 2021 based on the constructed optimal model, and spatial correlation analyses are performed between the fishing effort estimated in hours based on AIS data and in hooks based on fishing logbook data, by month. The results showed (1) the recognition accuracy of the threshold screening method is 89.9%, the recognition accuracy of the BP neural network classification model is 95.11%, the kappa coefficient is 0.51, the recognition accuracy of the SVM classification model is 95.74% and the kappa coefficient is 0.52; (2) in comparison, the SVM classification model performs better than the other two status identification methods for tuna longliners; and (3) the correlation coefficients between the two types of effort of VESSEL A were greater than 0.79 on all fishing months, indicating that there was no significant difference in the spatial and temporal distribution between the two types of effort. This study suggests that the SVM model can be used to identify the status and estimate the fishing effort of longliners.

Machine learning assisted strip waveguide Bragg gratings design for refractive index-based biosensing applications

Waveguide Bragg grating (WBG) biosensors have attracted significant interest because of their high sensitivity, immunity to electromagnetic signals, and capability for lab-on-chip applications. Designing a WBG biosensor requires solving Maxwell’s equations for the optical waveguide geometries numerically. This process is resource-intensive and hinders quick optimization of the waveguides for sensing. In this work, we explore the design of a silicon strip WBG structure for biosensing using machine learning. We develop an artificial neural network (ANN) classification model for mode classification and a regression model for predicting the effective refractive index (neff) for a given strip waveguide. We utilize results from finite element method and Transfer Matrix method for strip WBG structures to train a multi-output ANN regression model for quickly predicting sensing parameters. The mode classification model provides an accuracy of over 99%, and the regression model predicts the neff with a mean absolute error (MAE) of 0.5%. The multi-output regression model predicts the sensitivity with 0.5% MAE and quality factor and maximum reflectivity with less than 3% MAE. ANN model training requires a few minutes of computational time and reduces the dependency on computationally expensive resources for optimization of the sensor design, thereby fast-tracking the biosensor design cycle.

Revealing the failure mechanism of high strength seawater coral aggregate reinforced concrete slabs based on acoustic emission technology: Parameter analysis and DBN-BPNN classification

In this study, high strength seawater coral aggregate concrete (SCAC) with a compressive strength of 80 MPa was prepared, and seawater coral aggregate reinforced concrete slab (SCARCS) was designed with epoxy-coated steel bars. Combining with the AE technology, monotonic flexural static tests were conducted on SCARCS specimens with three varying reinforcement ratios (i.e., 0.85%, 1.13%, and 1.41%). Results showed that the increasing the reinforcement ratio led to an increase in peak bearing capacity and a decrease in center deflections at the peak. The loading process can be divided into four stages based on the varying magnitudes of the b-value and Ib-value. The primary mode of failure was tensile microcracks at various stages, and the ratio of shear microcracks reached its maximum value in the final stage of loading. Besides, based on the signal intensity analysis, it can be observed that the scale and complexity of cracks propagation inside the SCARCS specimen increased as the reinforcement ratio rose. In addition, back propagation neural network (BPNN) classification model optimized by deep belief network (DBN) algorithm was proposed to predict the damage degree of SCARCS during loading, and the average prediction accuracy of the proposed model was shown to be over 89%. The classification outcomes demonstrated that with an increase in the reinforcement ratio of SCARCS, the proportion of AE signals in Cluster 3 increased, while the proportion of AE signals in Cluster 2 dropped. Notably, both the average frequency and energy presented an ascending trend. The findings of this study provided valuable insights for designing and assessing the safety of SCAC structures for essential island and reef engineering infrastructure.

A Deep Quantum Convolutional Neural Network Based Facial Expression Recognition For Mental Health Analysis.

The purpose of this work is to analyze how new technologies can enhance clinical practice while also examining the physical traits of emotional expressiveness of face expression in a number of psychiatric illnesses. Hence, in this work, an automatic facial expression recognition system has been proposed that analyzes static, sequential, or video facial images from medical healthcare data to detect emotions in people's facial regions. The proposed method has been implemented in five steps. The first step is image preprocessing, where a facial region of interest has been segmented from the input image. The second component includes a classical deep feature representation and the quantum part that involves successive sets of quantum convolutional layers followed by random quantum variational circuits for feature learning. Here, the proposed system has attained a faster training approach using the proposed quantum convolutional neural network approach that takes [Formula: see text] time. In contrast, the classical convolutional neural network models have [Formula: see text] time. Additionally, some performance improvement techniques, such as image augmentation, fine-tuning, matrix normalization, and transfer learning methods, have been applied to the recognition system. Finally, the scores due to classical and quantum deep learning models are fused to improve the performance of the proposed method. Extensive experimentation with Karolinska-directed emotional faces (KDEF), Static Facial Expressions in the Wild (SFEW 2.0), and Facial Expression Recognition 2013 (FER-2013) benchmark databases and compared with other state-of-the-art methods that show the improvement of the proposed system.

A Method of Body Condition Scoring for Dairy Cows Based on Lightweight Convolution Neural Network

Highlights A cow body condition scoring model (ShuffleNet-SEH) based on a lightweight convolutional neural network using ShuffleNetV2 was proposed. Introducing the SE attention mechanism and H-Swish activation function enhances the feature extraction capability of the model. Achieved over 98% classification accuracy rate on the cow datasets. The proposed method can be applied to mobile devices. Abstract. Body condition scoring is an essential tool for nutrition management in large-scale dairy farming. The traditional method of manual scoring is inefficient and prone to subjective error. With the advancement of deep learning algorithms and machine vision technology, automatic scoring methods have been proposed. In response to the current problems of large parameter sizes and poor robustness in deep learning models, this study proposes a cow body condition scoring model (ShuffleNet-SEH) based on a lightweight convolutional neural network using ShuffleNetV2 as a foundation. Firstly, tail area images of cows were collected to construct a diverse dataset that includes complex variations such as motion blur and different levels of illumination. Secondly, the ShuffleNet-SEH model was developed by integrating a Squeeze-and-Excitation Networks (SENet) module into the main branch of ShuffleNetV2 after 1×1 convolution to enhance the representation ability of the network and improve its accuracy in identifying the cow body condition. Additionally, the nonlinear activation function ReLU in SENet was replaced by H-Swish to reduce the computational data required for the model on mobile devices, making it suitable for quantization operations. ShuffleNet-SEH was comprehensively evaluated, including ablation experiments, confusion matrix analysis, and feature map analysis. Multiple complex test sets were constructed for validation purposes. Throughout these evaluations, ShuffleNet-SEH exhibited remarkable performance. Furthermore, the efficacy and interpretability of ShuffleNet-SEH in cow body condition scoring tasks were further substantiated through meticulous feature map analysis. The findings underscore the robustness and reliability of ShuffleNet-SEH across diverse experimental scenarios and assessments. In the constructed motion blur test set, ShuffleNet-SEH demonstrated outstanding performance. Specifically, it achieved an accuracy rate of 98.2%, a precision rate of 98.5%, and a recall rate of 98.3%. These results represent significant improvements of 2.3%, 2.4%, and 3.2%, respectively, compared to the performance of the ShuffleNetV2 model. In addition, it is noteworthy that the ShuffleNet-SEH model has a size of approximately 5.64 MB, while the original ShuffleNet model has a size of 4.97 MB. This indicates a modest increase in model size of approximately 13.5% with the incorporation of the ShuffleNet-SEH enhancements. Moreover, superior overall performance was displayed by ShuffleNet-SEH model compared to mainstream convolutional neural network classification models, including MobileNetV2, EfficientNetV1, and ConvNeXt. The success of ShuffleNet-SEH in accurately identifying cow body condition while maintaining a lightweight architecture makes it suitable for deployment on mobile devices and, thus, has the potential to promote the commercialization of cow body condition scoring. Keywords: Attention mechanism, Body condition score, Deep learning, ShuffleNetV2.