This study was conducted in December 2024 at the Beef Cattle Pen, Livestock Development Center, Faculty of Animal Science, Universitas Gadjah Mada, Yogyakarta, Indonesia. The research aimed to develop an automated BCS assessment system by employing a YOLOv10-based deep learning architecture enhanced with CBAM and SPPFCSPC.

For image acquisition, a Hikvision 4MP AI CCTV camera was used, offering high-resolution video streaming with built-in artificial intelligence (AI) capabilities to enhance object detection. The camera was strategically mounted at a fixed height and angle to ensure consistent image capture for all cattle. Lighting conditions were standardized to minimize variations in image quality, following imaging protocols outlined by Xiong et al. (2023) and Spoliansky et al. (2016) to ensure consistent and replicable visual data for livestock monitoring.

Data processing and model training were conducted using a high-performance computing system, equipped with an Intel Core i9-12900K processor, 64GB RAM, and an NVIDIA RTX 3090 GPU (24GB VRAM). This setup facilitated efficient data handling and deep learning computations.

The dataset comprised 225 high-resolution images of beef cattle, systematically collected through controlled sampling sessions at the Faculty of Animal Science, Universitas Gadjah Mada. As shown in Table 1, the data acquisition process was conducted across three independent sessions, with each session involving 15 cattle, and five images captured per individual, resulting in 75 images per session and a total of 225 images. To ensure consistency and minimize variability, images were captured from a standardized single-angle perspective, providing uniform visual representations for subsequent analysis.

The study involved a total of 15 female beef cattle, comprising 12 Peranakan Ongole (PO), 2 Simmental, and 1 Brahman Cross. These animals were purposefully selected to represent the genetic and physiological diversity commonly found in Indonesian beef cattle production systems. The exclusive use of females also aligns with the practical context of BCS monitoring, as body condition is particularly critical in managing reproduction and lactation performance in female cattle.

The sampled cattle exhibited heterogeneous physiological characteristics, encompassing variations in sex (male and female), age groups (young and mature), and reproductive status (pregnant, lactating, or non-lactating). The structured multi-session data acquisition approach facilitated a diverse representation of body conditions while maintaining environmental consistency, thereby enhancing dataset robustness for deep learning-based BCS estimation (Patterson et al., 2021). BCS assessment can be conducted using visual indicators alone or a combination of visual and tactile estimations of key skeletal structures to evaluate fat cover. Key anatomical regions for BCS evaluation include the backbone, hips, tailhead, pins, rump, thigh, short ribs, and long ribs, as shown in Figure 1.

These regions are critical visual and tactile points for determining fat distribution and overall body condition. Visual assessment evaluates fat thickness over the ribs, tailhead, and backbone, providing essential cues about the cattle’s nutritional and health status. Exclusion criteria involved images with motion blur, partial visibility, or obstructed views that could affect accurate annotation. Figure 2 labels each image based on BCS categories (1–5) by two independent expert evaluators from the Faculty of Animal Science, Universitas Gadjah Mada, with more than five years of experience in manual BCS scoring. Inter-annotator agreement was verified, and any discrepancies were resolved through discussion to ensure biologically valid ground truth. While the dataset consisted of over 400 annotated images, these were derived from 15 individual cattle. The limited number of unique animals is acknowledged as a constraint regarding biological variability and population representation.

Captured images were preprocessed by resizing them to 640 × 640 pixels to match the input requirements of the YOLO model. From Table 2, techniques improved the model’s robustness to lighting, orientation, and background conditions variations. The YOLOv10 model was trained using supervised learning with 432 labeled images. The model was optimized using the Adam optimizer, with a learning rate of 0.001 and a batch size of 16, trained over 100 epochs on an NVIDIA RTX 3090 GPU to accelerate computation. Adam was selected due to its adaptive learning rate and faster convergence properties compared to traditional optimizers such as SGD or RMSprop, making it particularly effective for training deep neural networks on relatively small and heterogeneous datasets. To increase data diversity and model robustness, the original 225 images were augmented using standard techniques (rotation, brightness, blur, flipping, etc.), resulting in 432 labeled images used for model training.

An enhanced YOLOv10n architecture was developed by integrating CBAM and SPPFCSPC to improve feature extraction, detection accuracy, and computational efficiency for automated BCS assessment. As illustrated in Figure 3, the model consists of three primary components: Backbone, Neck, and Head, each modified to optimize multi-scale feature representation and object detection performance.

The backbone is designed for feature extraction, incorporating convolutional layers, CSP blocks, SCDown, and depthwise separable convolutions to enhance computational efficiency. In the YOLOv10n + SPPFCSPC configuration (Figure 3c), SPPFCSPC is added to refine multi-scale feature aggregation further, improving context-aware encoding critical for accurate BCS classification. The neck is responsible for feature refinement and fusion, ensuring that spatial and contextual information is effectively utilized. In the YOLOv10n + CBAM model (Figure 3b), CBAM is integrated into the neck to enhance channel and spatial attention, improving detection precision by emphasizing fine-grained fat distribution differences.

The head performs final BCS classification and object detection using a dual-branch prediction mechanism. The One-to-Many Prediction Head evaluates multiple bounding boxes, increasing robustness in complex scenes, while the One-to-One Prediction Head ensures accurate and efficient classification with minimal redundancy. These modifications collectively enhance the accuracy, computational efficiency, and real-time applicability of BCS assessment, making the model well-suited for scalable, precision livestock management.

In this study, nine YOLOv10-based model configurations were developed and evaluated to explore the impact of attention mechanisms and pooling strategies across different architecture scales. These configurations consist of three baseline models: YOLOv10n, YOLOv10m, and YOLOv10x, representing small, medium, and large variants, respectively. Each baseline was further enhanced with either the CBAM or the SPPFCSPC, resulting in six additional variants. This structured approach allows for a comprehensive analysis of how model complexity and specific architectural enhancements influence detection accuracy, computational efficiency, and BCS classification performance.

The CBAM is an attention mechanism designed to enhance CNN performance by emphasizing relevant features. CBAM is a lightweight module that integrates seamlessly with CNN architectures. Inspired by various attention mechanisms like squeeze-and-excitation networks (SENet) from 2017, Residual Attention Network from 2017 (Wang et al., 2024), and spatial transformer network (STN) from 2015, CBAM contains two sub-modules, the channel attention module and the spatial attention module, sequentially within a single attention block.

In Figure 4, designed to highlight critical features within each channel of a feature map. It takes the feature map as input and computes channel attention through two aggregation operations: average pooling and max pooling. These results are combined and then passed through fully connected layers and a sigmoid activation function to produce a channel attention map, which weighs the original feature map.

The SPPFCSPC module (spatial pyramid pooling fast convolutional special pyramid pooling convolution) draws inspiration from previous modules and techniques aimed at enhancing object detection capabilities, particularly in detecting small targets and accelerating processing. Among these is SPPF, or spatial pyramid pooling fast, an optimized version of SPP designed for speed, introduced in YOLOv5(Jooshin et al., 2024).

In Figure 5, pooling is performed sequentially across four different window sizes, with each pooling result passed to the subsequent pooling layer. SPPFCSPC incorporates the principles of SPPF, forwarding the pooled outputs to the next layer to retain spatial information and enhance processing speed. By integrating cross stage partial connections (CSPC), the module reduces computational redundancy and optimizes gradient flow, which is crucial for deep network architectures (Wang et al., 2023).

The combination of SPPF and CSPC allows for efficient multi-scale feature extraction, ensuring better object detection performance, particularly for small-scale features relevant to BCS assessment. This refinement is particularly beneficial in scenarios where precise localization of anatomical features is required for accurate classification (Nagy et al., 2023).

The performance evaluation of each YOLOv10-based model was conducted using four main metrics: Precision, Recall, F1-Score, and Mean Average Precision (mAP). As defined in Equation (1), precision measures the proportion of true positive predictions out of all positive predictions made by the model, reflecting its accuracy in identifying relevant instances. Recall, shown in Equation (2), quantifies the proportion of actual positives that were correctly identified, indicating the model’s sensitivity. These two metrics are combined into the F1-Score (Equation 3), which represents the harmonic mean of Precision and Recall and is particularly useful when evaluating imbalanced datasets. Moreover, the classification performance across all classes is summarized using mean average precision (mAP), as shown in Equation (4).

Precision= TP/(TP + FP) (1)

Recall= TP/(TP + FN) (2)

F1-score= 2 · (Precision · Recall)/(Precision + Recall) (3)

To evaluate the computational performance of the proposed enhanced YOLOv10 model, three key metrics were analyzed: inference time (ms), number of parameters (million), and floating-point operations per second (FLOPS, GigaFLOPS/G). A lower inference time is crucial for real-time BCS assessment, ensuring efficient large-scale cattle monitoring.

While the incorporation of CBAM, and SPPFCSPC slightly increases computational overhead, optimizations such as depthwise separable convolutions and CSPC mitigate excessive latency, maintaining a balance between accuracy and efficiency. Research has shown that cross stage partial networks (CSPNet) effectively reduce redundancy while preserving high feature representation quality, making them suitable for lightweight yet accurate detection models (Wang et al., 2023).

The number of parameters and FLOPS indicate model complexity and computational cost, impacting real-time deployment feasibility. Although the proposed model introduces additional layers for enhanced feature extraction, leveraging attention-based mechanisms and efficient pooling strategies prevents excessive growth in computational demand. Studies on transformer-based architectures highlight the importance of optimizing attention mechanisms to maintain high accuracy with minimal processing overhead, ensuring real-time applicability.

The dataset utilized in this study was annotated using Roboflow, where each image was manually labeled to identify key regions of interest (ROI), ensuring precise detection and segmentation of cattle bodies. These annotations served as the ground truth for model training and evaluation.

Figure 6 presents examples of annotated images, demonstrating the labeling consistency across different BCS categories (1–5). The high-quality annotations contribute to the model’s ability to generalize well across diverse cattle conditions.

Figure 7 illustrates the normalized confusion matrices for nine different YOLOv10-based model configurations, each reflecting the model’s performance in classifying cattle across BCS categories 1 to 5.

Across the matrices, models such as YOLOv10x (Figure 7g) and YOLOv10n (Figure 7a) display stronger diagonals, suggesting more consistent and accurate predictions. In contrast, configurations involving single-module integrations (e.g., YOLOv10n + CBAM in Figure 7b or YOLOv10m + SPPFCSPC in Figure 7f) demonstrate relatively dispersed predictions, indicating potential confusion between adjacent BCS classes. The visual comparison underscores that deeper models or those with balanced attention and pooling mechanisms tend to yield improved category-specific accuracy, particularly in distinguishing subtle differences in intermediate BCS categories (BCS 2, 3, and 4).

A comparative analysis of different YOLOv10-based models for BCS assessment, focusing on accuracy, precision, recall, and F1 score, is presented in Table 3. The results indicate that YOLOv10x outperforms all other models, achieving the highest accuracy (88.2%), precision (88.2%), recall (83.6%), and F1 score (85.8%), demonstrating its superior capability in both classification accuracy and robustness. Among the YOLOv10n variants, the base YOLOv10n model (82.4% Accuracy, 80.9% F1 Score) performs better than YOLOv10n + CBAM (83.8% Accuracy, 69.5% F1 Score) and YOLOv10n + SPPFCSPC (71.2% Accuracy, 72.0% F1 Score).

For YOLOv10m models, YOLOv10m + CBAM (83.7% Accuracy, 78.5% F1 Score) slightly outperforms the base model (79.4% Accuracy, 79.2% F1 Score), while YOLOv10m + SPPFCSPC (75.4% Accuracy, 72.8% F1 Score) exhibits a decline in overall performance. This suggests that SPPFCSPC does not provide substantial performance gains in mid-sized YOLO architectures for BCS classification.

Additionally, YOLOv10x + CBAM (86.5% Accuracy, 79.7% F1 Score) and YOLOv10x + SPPFCSPC (84.0% Accuracy, 82.8% F1 Score) show that SPPFCSPC improves recall (81.7%) compared to CBAM (73.9%), but at the cost of lower overall accuracy. These findings emphasize the trade-off between feature extraction complexity and classification efficiency, where deeper models (YOLOv10x) perform better at multi-scale feature learning than their compact counterparts (YOLOv10n and YOLOv10m).

The YOLOv10x model demonstrated the highest performance in BCS classification, achieving an overall accuracy of 88.2% and an F1 Score of 85.8%.

Despite this superior accuracy, the model incurred a substantially higher computational cost, which may limit its applicability in resource-constrained environments. With an inference time of 5.7 ms, 31.59 million parameters, and 169.8 GFLOPS, YOLOv10x is computationally intensive and better suited for high-resource environments such as cloud-based systems or offline batch processing (Table 4).

In contrast, the YOLOv10n + CBAM model delivered a competitive accuracy (83.8%) with an F1 Score of 69.5%, while significantly reducing the computational burden. It maintained a minimal inference time of 0.9 ms, a lightweight architecture with 2.70 million parameters, and only 8.2 GFLOPS—the same as the baseline YOLOv10n. The integration of CBAM refined attention to spatial and channel-wise features, improving precision without sacrificing efficiency.