Adaptive Stochastic YOLO for Real-Time Meat Freshness Detection in Food Safety Inspection

Animal protein sources are essential for health, as they help fulfil dietary requirements needed to prevent and treat stunting [1-4]. Beef is one of the most commonly consumed animal protein sources [5]. Beef freshness serves as a key quality indicator because spoilage can negatively impact both human health and the meat processing industry [6, 7]. Therefore, the accurate and real-time detection of beef freshness is very important for not only the quality maintenance of the products but also the food safety and security engineering in the meat supply chain.

Timely, reliable detection of beef freshness is crucial for beef quality assurance and for food safety and security within the meat supply chain. If quality degradation goes undetected before distribution or consumption, the meat may become unsafe. The World Health Organization (WHO) reports that foodborne diseases affect about 866 million people and cause 1.52 million deaths each year, with children under five bearing a substantial portion of the burden [8]. Food safety is a key component of food security: unsafe food threatens public health, erodes consumer confidence, and disrupts trade and supply-chain reliability [8].

Microbiological contamination is still a critical problem in beef and beef products. A recent systematic review and meta-analysis showed a pooled prevalence of 9.73% for Salmonella in over 106,000 beef samples, with a higher rate of contamination in processed beef products than in raw beef [9]. These results indicate the necessity of rapid, accurate and non-destructive inspection systems that can help in early detection and prevent potentially unsafe meat from reaching the food supply chain.

From a safety and security engineering perspective, automated beef freshness inspection serves as a risk-control mechanism in food production and distribution systems. The proposed system supports food safety inspection by identifying visual signs of freshness degradation before meat products reach consumers. It also contributes to hazard prevention by reducing the possibility of spoiled or unsafe meat entering distribution channels. When integrated into digital monitoring platforms, automated inspection can support supply-chain monitoring, traceability, quality assurance, regulatory compliance, and operational decision-making in food processing, storage, retail, and distribution environments [10]. Thus, beef freshness detection is not merely an image analysis task, but an engineering problem related to the design of safer, more reliable, and more secure food inspection workflows.

To date, beef freshness is mainly assessed by manual inspection (visual appearance, odor, and texture), which is subjective, inconsistent, and not scalable for large volumes [11-13]. Chemical or thermal sensors offer alternatives but are costly and complex. Therefore, automatic and scalable solutions are needed for beef freshness assessment, particularly for beef products.

Rapid developments in deep learning and computer vision in the last decade have opened new opportunities for image-based (non-destructive) automated food quality inspection. Deep Learning-based approaches have shown strong potential in extracting visual information for freshness detection [11, 14-16]. Image-based beef inspection is appealing because it can be implemented using conventional cameras without specialised sensors, while supporting real-time processing. Yet visual inspection of beef appearance faces challenges due to lighting variations, capture angles, temperature- and time-induced color changes, and surface reflections.

Among single-stage object detectors, You Only Look Once (YOLO) was introduced by Redmon [17] and stands out for its speed and accuracy. The architecture of YOLO has been continuously improved from YOLOv1 to YOLOv11 [17-24]. The main advantage of YOLO is its capability to detect objects in images and videos while keeping high inference speed [17, 25, 26]. This efficiency makes it the first choice for real-time visual inspection in the food-safety field to detect beef freshness. The suitability of YOLO for safety-critical real-time systems has been further confirmed in non-food domains, where it has been successfully deployed for nighttime surveillance [27].

YOLOv11, though presented as the latest model with strong performance in controlled environments, is essentially deterministic: a given image yields a single prediction. Under non-ideal real-world conditions, such as changing lighting, extreme camera viewpoints, or reflections, this deterministic prediction can become unstable or inaccurate. The lack of a mechanism to stabilize predictions across visual variations represents a critical limitation, especially for food-safety applications that are highly sensitive to errors.

This work proposes Adaptive Stochastic YOLO (AS-YOLO) to address these limitations through a stochastic refinement approach that operates after the YOLO detection stage. AS-YOLO does not alter the internal architecture or training process of YOLO; instead, it functions as an inference-time post-detection refinement mechanism. The core idea is to generate multiple prediction hypotheses from a single YOLO output through controlled stochastic perturbations and refine them using clustering and confidence-weighted fusion to obtain more stable detection outputs.

This study introduces a Stochastic Refinement Layer (SRL) to implement the proposed approach as a post-processing module consisting of four main steps. First, confidence-score perturbation is applied to generate controlled variations in detection confidence. Second, Gaussian bounding-box perturbation introduces small-scale spatial variations around the predicted bounding boxes. Third, class-constrained Intersection over Union (IoU)-based clustering groups overlapping prediction hypotheses with the same predicted class based on their IoU values, allowing detections that represent the same object to be identified. Finally, confidence-weighted voting combines multiple hypotheses within each cluster into a single refined detection, in which hypotheses with higher confidence scores contribute more strongly to the final output. The approach is stochastic because it applies controlled perturbations to the detection outputs during inference. However, it does not implement standard Monte Carlo Dropout, because it does not perform stochastic forward passes through the full YOLO network. The stochastic process is applied only at the detection-output level, keeping the computation lightweight and preserving real-time feasibility. Therefore, the resulting spatial variation should be interpreted as an output-level bounding-box variance indicator rather than full epistemic uncertainty estimation.

The main contributions are delineated as follows:

• This study proposes the development of the SRL as a lightweight post-detection refinement framework that can be integrated with existing YOLO detectors without modifying the backbone, neck, or detection head. The proposed SRL operates entirely at the detection-output level and does not require additional training or fine-tuning, making it practical for real-time inspection scenarios where architectural modification and retraining are not always feasible.

• This study introduces a stochastic perturbation mechanism that combines confidence-score perturbation and Gaussian bounding-box perturbation. This mechanism generates multiple prediction hypotheses from a single YOLO inference, allowing the detection output to be refined without repeated network forward passes.

• This study employs class-constrained IoU-based clustering and confidence-weighted voting to aggregate multiple prediction hypotheses into a final refined detection. The process is designed to maintain class consistency and bounding-box stability, while the resulting spatial variation is treated as a bounding-box variance indicator rather than full epistemic uncertainty estimation.

• The AS-YOLO framework has been tested on three variants of YOLO, namely YOLOv5, YOLOv8, and YOLOv11, using a beef image dataset comprising fresh and non-fresh classes. Model performance evaluation was conducted using standard evaluation metrics for object detection, including mean Average Precision (mAP), precision, recall, and inference speed, to assess detection performance and real-time feasibility.

• This study provides preliminary external testing using independently collected primary images to examine the behavior of AS-YOLO beyond the internal dataset. This evaluation is used to assess detection stability under independent acquisition conditions, while avoiding overgeneralization beyond the tested data.

In summary, the proposed AS-YOLO framework contributes to real-time beef freshness inspection by offering a lightweight output-level refinement mechanism that supports detection stability and real-time feasibility. Its relevance to safety and security engineering lies in its potential application for automated food quality inspection, hazard prevention, and supply-chain monitoring, while broader validation on larger and more diverse real-world datasets remains necessary.

1.png

Figure 1. Proposed Adaptive Stochastic YOLO (AS-YOLO) framework with Stochastic Refinement Layer (SRL) applied as a post-detection inference refinement module

1. Stochastic Perturbation

The first stage generates multiple prediction variations through output-level stochastic perturbation. In this implementation, stochastic perturbation is applied to the confidence score of each YOLO detection output, not to the normalized bounding-box coordinates. Inspired by Monte Carlo Dropout [43], this study does not perform stochastic sampling across the entire YOLO network. Instead, dropout is applied only at the detection-output level to introduce controlled variation in the confidence score. Therefore, this stage is implemented as a dropout-based confidence-score perturbation.

By restricting the stochastic process to the detection-output level, SRL should not be interpreted as a full epistemic uncertainty estimation method. Unlike conventional Monte Carlo Dropout, SRL does not perform repeated stochastic forward passes through the YOLO backbone, neck, or detection head. Instead, it generates output-level candidate variations that are later combined with Gaussian bounding-box perturbation, IoU-based clustering, and confidence-weighted voting. This setup enables SRL to generate multiple prediction hypotheses for each YOLO detection without modifying or retraining the model.

2. Gaussian Perturbation

After the stochastic perturbation stage, Gaussian perturbation is applied to the normalized bounding-box coordinates to generate controlled spatial variations around the original YOLO detection output. While stochastic perturbation is applied to the confidence score, Gaussian perturbation is applied only to the bounding-box coordinates. This perturbation is intended to support the refinement process in handling small spatial variations in detection outputs, which may occur due to differences in lighting conditions, camera angles, surface reflections, and meat texture appearance [49]. The class label remains unchanged from the original YOLO prediction.

Unlike more complex probabilistic approaches, this study uses a fixed Gaussian noise scale in each experimental configuration. This provides a simple output-level mechanism for generating spatial candidate variations without modifying or retraining the YOLO architecture. The Gaussian perturbation does not aim to estimate full predictive uncertainty; instead, it produces candidate bounding boxes around the initial YOLO detection for subsequent clustering and aggregation.

The Gaussian perturbation complements the confidence-score stochastic perturbation stage by generating multiple output-level candidate hypotheses. These candidates are then grouped using class-constrained IoU-based clustering and aggregated through confidence-weighted voting. Therefore, Gaussian perturbation is used as part of the SRL post-detection refinement process to support more stable and consistent detection outputs.

3. IoU Clustering

The outputs from the perturbation stages are grouped using an IoU-based clustering mechanism. This process is conceptually similar to Non-Maximum Suppression (NMS) [50]. However, in SRL, IoU-based clustering is used to group perturbed detection candidates before final aggregation, rather than to directly suppress overlapping predictions.

In this implementation, clustering is constrained by both spatial overlap and class consistency. Two prediction candidates are placed into the same cluster only if their IoU value is greater than or equal to the predefined threshold and they have the same predicted class label. This class-consistency constraint prevents candidates from different classes from being merged into the same cluster, even when their bounding boxes overlap.

By grouping overlapping candidates with the same class label, this stage reduces redundant hypotheses and prepares a more consistent candidate set for the confidence-weighted voting stage.

4. Confidence-Weighted Voting

After each cluster is formed, SRL applies confidence-weighted voting to generate the final prediction. Each candidate in a cluster is assigned a weight based on its confidence score, so candidates with higher confidence contribute more strongly to the final bounding-box estimate [47]. The final bounding box is calculated as the confidence-weighted average of all bounding boxes in the cluster, while the final confidence score is determined as the average confidence score of the clustered candidates. The final class label is determined from the class labels within the cluster.

In addition to producing the final bounding box, confidence score, and class label, SRL computes the spatial variance of bounding boxes within the cluster. This variance is used as a bounding-box variance indicator to describe the internal consistency of the clustered candidates. A lower variance value indicates that the perturbed candidates are spatially more consistent, whereas a higher variance value indicates larger variation among candidates. In this study, the bounding-box variance indicator is not used as a prediction rejection threshold and should not be interpreted as a full uncertainty estimation measure.

3.3 Mathematical formulation

The mathematical formulation in the SRL is designed to refine YOLO detection candidates by generating output-level stochastic perturbations, grouping overlapping predictions, and aggregating the final detection results through confidence-weighted voting. In this implementation, SRL is applied after a single YOLO inference and does not require repeated forward passes through the YOLO backbone, neck, or detection head.

Let the i-th detection candidate output from YOLO be defined as in Eq. (1):

$d_i=\left(b_i, s_i, c_i\right)$          (1)

where, $b_i$ is bounding box, $s_i$ is the confidence score, and $c_i$ is the class label [17]. The bounding box is expressed as Eq. (2):

$\mathrm{b}_i=\left(x_i, y_i, w_i, h_i\right)$             (2)

where, $x_i, y_i$ are the bounding box's normalized center coordinates, and $w_i$ and $h_i$ are the bounding box's normalized width and height, respectively.

3.3.1 Stochastic output perturbation iterations

Initial YOLO predictions are filtered by applying the confidence threshold $s_i>s_{\text {th }}$, where $s_{\text {th }}$ is the detection confidence threshold used in the corresponding experimental configuration. For each retained YOLO detection output, SRL generates $K$ stochastic output-level candidates as formulated in Eq. (3):

$\mathrm{d}_i^{(k)}=\left(b_i^{(k)}, s_i^{(k)}, c_i\right), k=1,2, \ldots, K$           (3)

where, $\bar{b}_i^{(k)}$ denotes the perturbed bounding box, $\widetilde{s}_i^{(k)}$ denotes the perturbed confidence score, and $c_i$ remains unchanged from the original YOLO prediction. Thus, SRL perturbs only the bounding-box coordinates and confidence scores at the output level, while the predicted class label is preserved during the stochastic candidate generation process.

The number of stochastic perturbation iterations was set to K = 10. This value was selected to provide a sufficient number of perturbed detection candidates for class-constrained IoU-based clustering and confidence-weighted voting, while maintaining computational efficiency during inference. Prior studies on MC Dropout-based uncertainty estimation have shown that prediction distributions tend to stabilize after several stochastic passes, commonly within the range of 10-20 iterations, with diminishing performance gains beyond that range [51]. Although SRL is not implemented as full MC Dropout, this range provides a useful empirical reference for determining the number of stochastic samples. In SRL, stochasticity is applied only at the detection-output level rather than through repeated stochastic forward passes across the entire YOLO network. Therefore, K = 10 was considered sufficient to generate diverse candidate variations around the initial YOLO detections without introducing excessive computational overhead. A smaller number of iterations may limit candidate diversity for reliable clustering, whereas a substantially larger number may increase inference latency and reduce suitability for real-time food safety inspection. Thus, K = 10 was used as a fixed sampling setting to ensure a fair comparison across SRL configurations, while the ablation study focused on Gaussian perturbation scale, confidence-score dropout probability, IoU threshold, and confidence threshold.

3.3.2 Stochastic perturbation

In this subsection, stochastic perturbation refers to dropout-based perturbation applied only to the confidence score of each YOLO detection output. It is not applied to the normalized bounding-box coordinates, nor to the full YOLO network. The bounding-box coordinates are perturbed separately using Gaussian perturbation, while dropout is used to introduce controlled stochastic variation in the detection confidence score at the output level.

Let s_i denote the original confidence score of the i-th YOLO detection candidate. For the k-th stochastic perturbation iteration, the perturbed confidence score is formulated as Eq. (4):

$\tilde{s}_j^{(k)}=\operatorname{Dropout}\left(s_i ; p\right)$               (4)

where p denotes the dropout probability. Since dropout is applied only at the detection-output level, this process should not be interpreted as full MC Dropout uncertainty estimation or Bayesian inference across the entire YOLO network. Instead, it is used as a confidence-score stochastic perturbation mechanism to generate variations among detection candidates before class-constrained IoU-based clustering and confidence-weighted voting.

The initial SRL configuration used p = 0.3 as an aggressive confidence-score perturbation setting. This value was selected by referring to prior MC Dropout-based studies that commonly evaluate dropout probabilities within the range of 0.2–0.5 [43]. In addition, Stochastic-YOLO-related work has investigated dropout rates of 25%, 50%, and 75%, indicating that moderate dropout settings can provide a practical balance between stochasticity and prediction stability in YOLO-based object detection systems [46]. Based on these references, p = 0.3 was used as the initial setting to provide sufficient stochastic variation in confidence scores.

However, because SRL applies dropout only to detection confidence scores rather than to full network activations or bounding-box coordinates, the effect of p may differ from conventional MC Dropout. Therefore, the dropout probability was further examined through the ablation study. In the ablation configurations, confidence-score dropout was removed in the Gaussian-only setting (p = 0), while lower dropout probabilities (p = 0.05 and p = 0.02) were evaluated to analyze whether reduced confidence-score perturbation could improve detection stability. Thus, p = 0.3 represents the initial aggressive configuration, whereas the lower dropout probabilities are treated as ablation settings for evaluating the sensitivity of SRL to confidence-score perturbation.

3.3.3 Gaussian perturbation

Gaussian perturbation is applied to the normalized bounding-box coordinates to generate controlled spatial variations around the original YOLO detection output. Unlike stochastic perturbation in the previous subsection, which is applied only to the confidence score, Gaussian perturbation is applied only to the bounding-box coordinates. The class label remains unchanged from the original YOLO prediction.

Let b_i denote the normalized bounding box of the i-th YOLO detection candidate. For the k-th stochastic perturbation iteration, the perturbed bounding box is formulated as Eq. (5):

$\bar{b}_i^{(k)}=\operatorname{clip}\left(b_i+\epsilon_i^{(k)}, 0,1\right), \epsilon_i^{(k)} \sim N\left(0, \sigma^2 I\right)$           (5)

where, $\epsilon_i^{(k)}$ represents Gaussian noise, $\sigma$ is the perturbation scale, and $I$ denotes the identity matrix for the four boundingbox components $(x, y, w, h)$ The clipping operation constrains the perturbed normalized bounding-box coordinates to the range $[0,1]$.

This perturbation is intended to support the refinement process in handling small spatial variations in detection outputs, which may occur due to differences in lighting conditions, camera angles, surface reflections, and meat texture appearance [49]. The use of Gaussian distributions is also consistent with previous studies that model spatial variation in bounding-box predictions [52]. In addition, perturbing bounding-box predictions has been reported to improve the robustness and generalization of object detection models [53].

The initial SRL configuration used σ = 0.05 as an aggressive spatial perturbation setting. This value was selected empirically to provide sufficient spatial variation around the initial YOLO detection while keeping the perturbed bounding boxes within reasonable normalized coordinate bounds. However, since excessive perturbation may shift detections away from the object region and reduce localization accuracy, smaller values of σ were further evaluated in the ablation study. Specifically, $\sigma=0.005$ was used in the Gaussian-only and moderate SRL configurations, while $\sigma=0.002$ was used in the conservative SRL configuration to analyze the effect of reduced spatial perturbation on detection stability.

3.3.4 Intersection over Union-based clustering

IoU for two bounding boxes $b_a$ and $b_b$, is computed using Eq. (6) [54]:

$\operatorname{IoU}\left(b_a, b_b\right)=\frac{\operatorname{Area}\left(b_a \cap b_b\right)}{\operatorname{Area}\left(b_a \cup b_b\right)}$         (6)

Two predictions are placed into the same cluster if they satisfy both the IoU overlap criterion and the class-consistency criterion $\operatorname{IoU}\left(b_a, b_b\right) \geq \tau_{\text {IoU }}$ and $c_a=c_b$, where $\tau_{\text {IoU }}$ is the IoU threshold, while $c_a$ and $c_b$ are the class labels of the two detection candidates. This means that two candidates are not grouped only because their bounding boxes overlap; they must also belong to the same predicted class.

This process is conceptually similar to non-maximum suppression, but in SRL it is used to group perturbed detection candidates before final aggregation rather than to directly suppress overlapping predictions [50]. The class-consistency constraint prevents candidates from different classes from being merged into the same cluster, even when their bounding boxes overlap.

The initial SRL configuration used $\tau_{\text {IoU }}=0.5$, while the ablation configurations used $\tau_{\text {IoU }}=0.6$. The higher IoU threshold in the ablation settings was used to enforce stricter spatial consistency among candidates within the same cluster, thereby reducing the possibility of grouping weakly overlapping predictions.

3.3.5 Weighted voting and final prediction

After the cluster $C$ is formed, the final prediction is computed using confidence-weighted voting. Each prediction in the cluster is assigned a weight based on its confidence score $w^{(n)}=s^{(n)}$, where $s^{(n)}$ denotes the confidence score of the nth prediction candidate in cluster $C$. The final bounding box is calculated as the confidence-weighted average of all bounding boxes in the cluster using Eq. (7):

$b_{\text {final }}=\frac{\sum_{n \in C} w^{(n)} b^{(n)}}{\sum_{n \in c^w} w^{(n)}+\varepsilon}$            (7)

where, $\epsilon=10^{-6}$ to prevent division by zero. This confidence-based aggregation is in line with the principle of weighted bounding box fusion used in the Weighted Boxes fusion method [47].

The final confidence score is determined as the average confidence score within the cluster using Eq. (8):

$S_{\text {final }}=\frac{1}{|C|} \sum_{n \in C} S^{(n)}$         (8)

The final class label is determined as the mode of the class IDs within the cluster according to Eq. (9):

$c_{\text {final }}=\operatorname{mode}\left(\left\{c^{(n)} \mid n \epsilon C\right\}\right)$           (9)

Since the clustering process is class-constrained, all candidates within the same cluster are required to have the same class label. Therefore, using the mode of the class IDs provides a formal representation of the final class assignment and remains consistent with the implementation.

3.3.6 Bounding-box variance indicator

In addition to generating the final bounding box and confidence score, SRL computes spatial variance as an indicator of bounding-box consistency within the cluster using Eq. (10):

$\operatorname{Var}(b)=\frac{\sum_{n \in C} w^{(n)}\left(b^{(n)}-b_{\text {final }}\right)^2}{\sum_{n \in C} w^{(n)}+\epsilon}$            (10)

A smaller variance value indicates that predictions within the cluster are more consistent; thus, the detection results are considered more stable. Conversely, a larger variance value indicates greater spatial variation among the perturbed candidates within the cluster. This value is therefore used as a bounding-box variance indicator, not as a full uncertainty estimation measure. The final AS-YOLO output is expressed as Eq. (11):

$D_{\text {final }}=\left(b_{\text {final }}, s_{\text {final }}, c_{\text {final }}, \operatorname{Var}_{(b)}\right)$.             (11)

In this implementation, $Var_{(b)}$ is used as an indicator of internal consistency, not as a threshold to reject predictions. Therefore, SRL should be understood as an output-level stochastic post-detection refinement module that aims to improve the stability and consistency of YOLO detection outputs without modifying or retraining the YOLO architecture.

This study concludes that AS-YOLO can be used as a lightweight post-detection refinement framework for YOLO-based beef freshness detection. The proposed SRL operates only at the detection-output level and does not modify the backbone, neck, or detection head of the YOLO models. It also does not require retraining. The refinement process is performed through confidence-score perturbation, Gaussian bounding-box perturbation, class-constrained IoU-based clustering, and confidence-weighted voting. Therefore, AS-YOLO should be interpreted as an output-level refinement approach for stabilizing existing YOLO predictions, not as a replacement for the detector architecture.

The experimental results show that the selected conservative SRL configuration, namely C4, maintains strong detection performance on the internal testing dataset. All baseline YOLO and AS-YOLO variants achieved 100.00% mAP@0.5 and 100.00% recall on the internal testing dataset. For the stricter localization metric, AS-YOLOv5, AS-YOLOv8, and AS-YOLOv11 achieved mAP@0.5:0.95 values of 96.70%, 96.82%, and 97.52%, respectively. These results indicate that the proposed refinement process can preserve class prediction and bounding-box consistency under the tested internal data conditions. However, because the paired bootstrap and McNemar tests did not show statistically significant differences between the baseline YOLO models and their AS-YOLO variants, the results should not be interpreted as evidence of statistically significant performance improvement. Instead, the main conclusion is that AS-YOLO maintains stable detection behavior while preserving the high performance of the baseline detectors.

The preliminary external testing using independently collected primary images further shows that the selected C4 configuration can maintain comparable detection behavior outside the internal dataset. However, the metric differences on the primary dataset were also not statistically significant. Therefore, the primary-data experiment provides an initial indication of detection stability under independent acquisition conditions, but it does not represent full external validation. This interpretation is important because real-world food safety inspection requires robustness across different camera devices, lighting conditions, reflections, backgrounds, meat storage conditions, and market or industrial environments.

In terms of practical relevance, AS-YOLO supports image-based beef freshness inspection as part of food safety monitoring and supply-chain quality control. The method is relevant to safety and security engineering because it can assist in preventing unsafe meat distribution through fast visual inspection, while maintaining real-time feasibility. All AS-YOLO variants achieved more than 30 FPS on both internal and primary testing datasets. On the internal testing dataset, AS-YOLOv5, AS-YOLOv8, and AS-YOLOv11 achieved 57.50 FPS, 54.54 FPS, and 52.36 FPS, respectively. On the primary dataset, the corresponding models achieved 45.06 FPS, 33.00 FPS, and 42.71 FPS. These results indicate that the additional refinement process introduces acceptable computational overhead for real-time inspection scenarios.

This study has several limitations. The internal evaluation was conducted on a two-class beef freshness dataset, and the independently collected primary dataset was limited in size. The proposed SRL also provides output-level stochastic refinement and spatial stability analysis, not full epistemic uncertainty estimation of the YOLO network. Future research should validate AS-YOLO using larger external datasets collected from real market and industrial environments, with different camera devices, lighting conditions, backgrounds, reflections, and storage durations. Further studies should also include additional meat types, such as chicken, fish, and goat meat, as well as more detailed freshness categories, including fresh, semi-fresh, and spoiled. Repeated-run experiments and broader statistical testing on aggregate detection metrics are also needed to further verify the robustness and generalizability of the proposed refinement framework.