Deep Learning-Based Meat Freshness Detection with Segmentation and OOD-Aware Classification

Hutama Arif Bramantyo; Mukarram Ali Faridi; Rui Chen; Clarissa Harris; Yin Sun

cs.LGFeb 27, 2026

Deep Learning-Based Meat Freshness Detection with Segmentation and OOD-Aware Classification

Authors: Hutama Arif Bramantyo, Mukarram Ali Faridi, Rui Chen, Clarissa Harris, Yin Sun

Organizations: Electrical & Computer Engineering Department, College of Engineering, Auburn University, Auburn, AL 36849 USA · Department of Agricultural and Environmental Sciences, College of Agriculture, Environment & Nutrition Sciences, Tuskegee University, Tuskegee, AL 36088 USA

Abstract

In this study, we present a meat freshness classification framework from Red-Green-Blue (RGB) images that supports both packaged and unpackaged meat datasets. The system classifies four in-distribution (ID) meat classes and uses an out-of-distribution (OOD)-aware abstention mechanism that flags low-confidence samples as No Result. The pipeline combines U-Net-based segmentation with deep feature classifiers. Segmentation is used as a preprocessing step to isolate the meat region and reduce background, producing more consistent inputs for classification. The segmentation module achieved an Intersection over Union (IoU) of 75% and a Dice coefficient of 82%, producing standardized inputs for the classification stage. For classification, we benchmark five backbones: Residual Network-50 (ResNet-50), Vision Transformer-Base/16 (ViT-B/16), Swin Transformer-Tiny (Swin-T), EfficientNet-B0, and MobileNetV3-Small. We use nested 5x3 cross-validation (CV) for model selection and hyperparameter tuning. On the held-out ID test set, EfficientNet-B0 achieves the highest accuracy (98.10%), followed by ResNet-50 and MobileNetV3-Small (both 97.63%) and Swin-T (97.51%), while ViT-B/16 is lower (94.42%). We additionally evaluate OOD scoring and thresholding using standard OOD metrics and sensitivity analysis over the abstention threshold. Finally, we report on-device latency using TensorFlow Lite (TFLite) on a smartphone, highlighting practical accuracy-latency trade-offs for future deployment.

Figures & tables

Fig. 1: Experiment setup for packaged meat image collection.

Fig. 2: Representative examples of the No Result (irrelevant/OOD) set, including empty trays, plates, and background-only images.

Category	Train	Val	Test	Total	Share (%)
Packaged Fresh Meat	656	138	146	940	15.0
Packaged Spoiled Meat	805	170	180	1,155	18.5
Unpackaged Fresh Meat	1,159	246	256	1,661	26.6
Unpackaged Spoiled Meat	1,187	252	261	1,700	27.2
Total	3,807	806	843	5,456	100.0

TABLE I: Per-class distribution after cleaning and stratified 70/15/15 splitting for the four in-distribution freshness classes. Unpackaged sources: [ 32 , 33 , 34 ] ; Packaged: this study (Apr 2024).

Hyperparameter	Search Range / Setting
Head learning rate (Inner-1)	$1{\times}10^{-3}$ – $3{\times}10^{-3}$
Backbone learning rate (Inner-2)	$1{\times}10^{-5}$ – $3{\times}10^{-4}$
Weight decay (Inner-1)	$1{\times}10^{-4}$ – $1{\times}10^{-1}$
Label smoothing $\alpha$	$\{0.0,\,0.1\}$
MixUp $\alpha$	$\{0.0,\,0.2\}$
Batch size	32

TABLE II: Hyperparameter search space explored during inner-loop tuning under the nested 5

\times

3 CV framework (classification).

Backbone	IoU	Dice	Precision	Recall	Pixel Acc.
ResNet-18	0.7211 [0.703, 0.739]	0.8021 [0.786, 0.817]	0.7780 [0.761, 0.794]	0.8727 [0.857, 0.888]	0.8569 [0.846, 0.868]
ResNet-34	0.7494 [0.731, 0.767]	0.8188 [0.802, 0.835]	0.8186 [0.802, 0.835]	0.8490 [0.832, 0.866]	0.8777 [0.867, 0.889]
ResNet-50	0.7320 [0.713, 0.750]	0.8057 [0.789, 0.822]	0.8103 [0.793, 0.826]	0.8419 [0.825, 0.859]	0.8709 [0.860, 0.882]

TABLE III: Global segmentation performance of U-Net backbones with 95% confidence intervals.

Backbone	Packaged Fresh	Packaged Spoiled	Unpackaged Fresh	Unpackaged Spoiled
ResNet-18	0.8726 [0.842, 0.900]	0.7749 [0.741, 0.808]	0.6693 [0.639, 0.699]	0.6501 [0.612, 0.687]
ResNet-34	0.8888 [0.859, 0.915]	0.8072 [0.771, 0.842]	0.6898 [0.659, 0.722]	0.6900 [0.653, 0.728]
ResNet-50	0.8900 [0.858, 0.918]	0.7997 [0.763, 0.834]	0.6738 [0.643, 0.704]	0.6541 [0.615, 0.693]

TABLE IV: Per-class IoU performance with 95% confidence intervals.

Model	Acc (%)	F1 (%)	Prec. (%)	Rec. (%)
ResNet-50	98.47 $\pm$ 0.37	98.47 $\pm$ 0.37	98.51 $\pm$ 0.35	98.47 $\pm$ 0.37
Swin-T	97.75 $\pm$ 0.64	97.75 $\pm$ 0.64	97.80 $\pm$ 0.62	97.75 $\pm$ 0.64
ViT-B/16	94.16 $\pm$ 1.03	94.13 $\pm$ 1.05	94.67 $\pm$ 0.76	94.16 $\pm$ 1.03
EfficientNet-B0	97.88 $\pm$ 0.72	97.88 $\pm$ 0.72	97.96 $\pm$ 0.67	97.88 $\pm$ 0.72
MobileNetV3-Small	96.64 $\pm$ 0.93	96.63 $\pm$ 0.93	96.91 $\pm$ 0.79	96.64 $\pm$ 0.93

TABLE V: Nested CV stability (outer-fold mean

\pm

SD).

Model	Acc. (Mean $\pm$ SD, %)	Head LR	Backbone LR	Weight Decay	Label Smooth	Mixup $\alpha$	Best Epoch
ResNet-50	98.47 $\pm$ 0.37	0.003	1e $-$ 4	0.0005	0.0	0.0	19–30
Swin-T	97.75 $\pm$ 0.64	0.0015	7e $-$ 5	0.1	0.0	0.0	13–30
ViT-B/16	94.16 $\pm$ 1.03	0.001	5e $-$ 5	0.05	0.1	0.0	13–30
EfficientNet-B0	97.88 $\pm$ 0.72	0.001	0.0003	0.0001	0.1	0.0	17–30
MobileNetV3-Small	96.64 $\pm$ 0.93	0.001	0.0003	0.0005	0.0	0.2	17–30

TABLE VI: Nested cross-validation (5 outer

\times

3 inner) results and best hyperparameters per model.

Outer Fold	ResNet-50 (Acc %)	Swin-T (Acc %)	ViT-B/16 (Acc %)	EfficientNet-B0 (Acc %)	MobileNetV3-Small (Acc %)
Fold 1	98.12	97.18	93.05	97.92	95.09
Fold 2	98.36	97.61	94.92	97.64	97.26
Fold 3	98.89	98.24	95.21	98.11	96.69
Fold 4	98.53	97.82	93.84	96.88	97.45
Fold 5	98.45	97.89	93.76	98.87	96.69
Mean $\pm$ SD	98.47 $\pm$ 0.37	97.75 $\pm$ 0.64	94.16 $\pm$ 1.03	97.88 $\pm$ 0.72	96.64 $\pm$ 0.93

TABLE VII: Outer-fold accuracies (%) under nested 5

\times

3 CV.

Model	Train	Val	Test (ID)
ResNet-50	99.9	98.5	97.63
Swin-T	99.7	98.6	97.51
ViT-B/16	99.1	94.2	94.42
EfficientNet-B0	99.9	97.2	98.10
MobileNetV3-Small	99.9	97.5	97.63

TABLE VIII: Retrain results on the held-out ID test set.

Fig. 3: Training vs validation accuracy curves for five models during retraining with OOD-enabled inference.

Class	ResNet-50			Swin-T			ViT-B/16			EfficientNet-B0			MobileNetV3-Small
Class	Prec.	Rec.	F1	Prec.	Rec.	F1	Prec.	Rec.	F1	Prec.	Rec.	F1	Prec.	Rec.	F1
Packaged Fresh	0.986	0.952	0.969	0.935	0.993	0.963	0.971	0.911	0.940	0.973	0.993	0.983	0.954	0.986	0.970
Packaged Spoiled	0.962	0.989	0.975	0.994	0.944	0.969	0.931	0.978	0.954	0.994	0.978	0.986	0.989	0.961	0.975
Unpackaged Fresh	0.959	1.000	0.979	0.992	0.969	0.980	0.895	1.000	0.945	0.959	1.000	0.979	0.959	1.000	0.979
Unpackaged Spoiled	1.000	0.958	0.978	0.970	0.992	0.981	1.000	0.885	0.939	1.000	0.958	0.978	1.000	0.958	0.978
Macro Avg	0.977	0.975	0.975	0.973	0.975	0.973	0.949	0.943	0.944	0.982	0.982	0.982	0.975	0.976	0.975

TABLE IX: Per-class performance of five models on the held-out ID test set (

N=843

Panel of Figure 13 — Fig. 4: Confusion matrices on the held-out ID test set ( $N=843$ ) for the four in-distribution classes (Packaged Fresh, Packaged Spoiled, Unpackaged Fresh, Unpackaged Spoiled). Errors mostly occur within the same packaging type (fresh vs. spoiled), which is the hardest visual boundary.

Panel of Figure 14 — Fig. 5: Representative failure cases: (i) glare on packaged samples, (ii) occlusion/clutter, and (iii) borderline freshness with subtle cues.

Model	Method	AUROC	AUPR	FPR@95TPR
ResNet-50	MSP	0.927	0.911	0.242
	ENERGY	0.939	0.941	0.254
	ODIN	0.930	0.927	0.281
Swin-T	MSP	0.981	0.992	0.0250
	Energy	0.996	0.998	0.0025
	ODIN	0.997	0.997	0.0025

TABLE X: OOD detection results across scoring methods on the OOD benchmark.

Fig. 6: Threshold sweep on the in-distribution (4-class) set: $\tau$ vs. coverage (fraction of samples kept).

Fig. 7: Threshold sweep on the in-distribution (4-class) set: $\tau$ vs. rejection rate (fraction of samples abstained).

Pair	$\mathbf{n_{11}}$	$\mathbf{n_{10}}$	$\mathbf{n_{01}}$	$\mathbf{n_{00}}$	$\bm{\chi^{2}}$	$\mathbf{p}$ -value	$\bm{\Delta}\,\mathrm{acc}$ (A–B)	95% CI
ResNet50 (A) vs ViT (B)	788	35	8	12	16.331	$5.32\times 10^{-5}$	$+0.0320$	$[0.0169,\ 0.0471]$
ViT (A) vs Swin-T (B)	778	18	44	3	10.512	$0.0012$	$-0.0308$	$[-0.0490,\ -0.0127]$
ViT (A) vs EfficientNet-B0 (B)	790	6	37	10	21.605	$3.35\times 10^{-6}$	$-0.0368$	$[-0.0518,\ -0.0217]$
ViT (A) vs MobileNetV3-Small (B)	780	16	43	4	11.932	$0.0005$	$-0.0320$	$[-0.0498,\ -0.0143]$
Swin-T (A) vs MobileNetV3-Small (B)	815	7	8	13	0.0167	$0.8973$	$-0.0012$	$[-0.0102,\ 0.0078]$

TABLE XI: McNemar paired contingency tables and confidence intervals for pairwise comparison of test accuracy on the held-out ID test set (

N=843

). McNemar

\chi^{2}

statistics use the continuity-corrected form

(|n_{10}-n_{01}|-0.5)^{2}/(n_{10}+n_{01})

(\mathrm{df}=1)

Class	ResNet-50	Swin-T	ViT-B/16	EfficientNet-B0	MobileNetV3-Small
Packaged Fresh	0.950 (0.921–0.974)	0.944 (0.916–0.969)	0.889 (0.847–0.925)	0.946 (0.918–0.970)	0.932 (0.899–0.960)
Packaged Spoiled	0.958 (0.933–0.978)	0.951 (0.927–0.973)	0.921 (0.891–0.947)	0.955 (0.932–0.975)	0.939 (0.911–0.964)
Unpackaged Fresh	0.982 (0.970–0.993)	0.982 (0.970–0.993)	0.936 (0.913–0.956)	0.975 (0.960–0.988)	0.961 (0.943–0.976)
Unpackaged Spoiled	0.983 (0.970–0.993)	0.983 (0.971–0.993)	0.928 (0.903–0.951)	0.974 (0.959–0.987)	0.958 (0.939–0.975)
Macro Avg	0.968 (0.954–0.980)	0.965 (0.951–0.978)	0.919 (0.899–0.938)	0.963 (0.948–0.976)	0.947 (0.930–0.963)

TABLE XII: Per-class F1-score with 95% bootstrap confidence intervals (percentile,

B=4000

) on the held-out ID test set (

N=843

Study	Classes	Dataset Size	Reported Metric
Ulucan et al. (2019) [ 17 ]	2	1,896	99.6% (Accuracy)
NUS Project (2023) [ 18 ]	3	2,268	93.1% (Accuracy)
Abd Elfattah et al. (2025) [ 19 ]	3	2,266	98.5% (Accuracy)
This Work (2026)	4 (ID)	6,256	98.10% (EfficientNet-B0), 97.63% (ResNet-50 / MobileNetV3-Small), 97.51% (Swin-T)

TABLE XIII: Quantitative comparison of prior RGB-based meat-freshness studies and this work. Results from this study are reported on the held-out ID test set. OOD performance is evaluated separately (Section IV-D ).

Model	Params (M)	Latency (ms)	Device / Runtime	Accuracy (%)	Train Time (min/epoch)
ResNet–50	25.6	76.92	Galaxy A55, Android 16, TFLite Benchmark	97.63	1.10
EfficientNet–B0	5.3	17.21	Galaxy A55, Android 16, TFLite Benchmark	98.10	0.76
MobileNetV3–Small	2.5–2.9	6.36	Galaxy A55, Android 16, TFLite Benchmark	97.63	0.74
Swin–T (Tiny)	27.5–28	108.17	Galaxy A55, Android 16, TFLite Benchmark	97.51	0.78
ViT–B/16	86	462.75	Galaxy A55, Android 16, TFLite Benchmark	94.42	0.84

TABLE XIV: Computational efficiency. Test accuracy and training time were measured on our lab GPU setup (RTX 6000 Ada), while mobile latency was benchmarked by us on a Samsung Galaxy A55 (SM-A556E), Android 16, using the TensorFlow Lite Benchmark APK (this work). Accuracy corresponds to the held-out ID test accuracy (Table VIII ).

Classification

Classification maps observations to discrete labels, including familiar, unknown, hierarchical, or multiple categories.

See all 410 papers

MyoVision: A Mobile Research Tool and NEATBoost-Attention Ensemble Framework for Real Time Chicken Breast Myopathy Detection
Apr 15, 2026Chaitanya Pallerla, Siavash Mahmoudi, Dongyi WangEnsemble Learning Computational Imaging
Department of Biological and Agricultural Engineering, University of Arkansas, Fayetteville, AR · Department of Food Science, University of Arkansas, Fayetteville, AR
Prototype Fusion: A Training-Free Multi-Layer Approach to OOD Detection
Mar 24, 2026Shreen Gul, Mohamed Elmahallawy, Ardhendu Tripathy +1OOD Detection
Missouri University of Science and Technology, Rolla, MO 65401, USA · Washington State University, Richland, WA 99354, USA
Domain-Aware Lightweight Spectral-Grouped Convolutions for Hyperspectral Fish Freshness Classification
Aug 12, 2026Kazi Nabiul Alam, Pooneh Bagheri Zadeh, Akbar Sheikh-AkbariRepresentation Learning Hyperspectral Image Classification
School of Built Environment, Engineering, and Computing, Leeds Beckett University. Leeds LS6 3QS, United Kingdom.

Apr 15, 2026J/K move · Enter open · S save

MyoVision: A Mobile Research Tool and NEATBoost-Attention Ensemble Framework for Real Time Chicken Breast Myopathy Detection

Chaitanya Pallerla, Siavash Mahmoudi, Dongyi Wang

Department of Biological and Agricultural Engineering, University of Arkansas, Fayetteville, AR · Department of Food Science, University of Arkansas, Fayetteville, AR

Ensemble Learning Computational Imaging

Open paper AskPDF arXiv

Woody Breast (WB) and Spaghetti Meat (SM) myopathies significantly impact poultry meat quality, yet current detection methods rely either on subjective manual evaluation or costly laboratory-grade imaging systems. We address the problem of low-cost, non-destructive multi-class myopathy classification using consumer smartphones. MyoVision is introduced as a mobile transillumination imaging framework in which 14-bit RAW images are captured and structural texture descriptors indicative of internal tissue abnormalities are extracted. To classify three categories (Normal, Woody Breast, Spaghetti Meat), we propose a NEATBoost-Attention Ensemble model, which is a neuroevolution-optimized weighted fusion of LightGBM and attention-based MLP models. Hyperparameters are automatically discovered using NeuroEvolution of Augmenting Topologies (NEAT), eliminating manual tuning and enabling architecture diversity for small tabular datasets. On a dataset of 336 fillets collected from a commercial processing facility, our method achieves 82.4% test accuracy (F1 = 0.83), outperforming conventional machine learning and deep learning baselines and matching performance reported by hyperspectral imaging systems costing orders of magnitude more. Beyond classification performance, MyoVision establishes a reproducible mobile RGB-D acquisition pipeline for multimodal meat quality research, demonstrating that consumer-grade imaging can support scalable internal tissue assessment.

Abstract

Classification

Explore similar work

MyoVision: A Mobile Research Tool and NEATBoost-Attention Ensemble Framework for Real Time Chicken Breast Myopathy Detection

Prototype Fusion: A Training-Free Multi-Layer Approach to OOD Detection

Domain-Aware Lightweight Spectral-Grouped Convolutions for Hyperspectral Fish Freshness Classification