Visual Anomaly Detection (VAD) is a critical task for many applications including industrial inspection and healthcare. While VAD has been extensively studied, two key challenges remain largely unaddressed in conjunction: edge deployment, where computational resources are severely constrained, and continual learning, where models must adapt to evolving data distributions without forgetting previously acquired knowledge. Our benchmark provides guidance for the selection of the optimal backbone and VAD method under joint efficiency and adaptability constraints, characterizing the trade-offs between memory footprint, inference cost, and detection performance. Studying these challenges in isolation is insufficient, as methods designed for one setting make assumptions that break down when the other constraint is simultaneously imposed. In this work, we propose the first comprehensive benchmark for VAD on the edge in the continual learning scenario, evaluating seven VAD models across three lightweight backbone architectures. Furthermore, we propose Tiny-Dinomaly, a lightweight adaptation of the Dinomaly model built on the DINO foundation model that achieves a 11× smaller memory footprint and 20× lower computational cost while improving localization (Pixel F1) by 5 percentage points. Finally, we introduce targeted modifications to PatchCore and PaDiM to improve their efficiency in the continual learning setting.
Figures & tables
Figure 1: Continual Learning setting for the VAD. Each task introduces a new item, the VAD model must identify anomalous products at the image level and localize defects at the pixel level for newly introduced items, while maintaining strong performance on previously learned categories.
Figure 2: Performance (Pixel F1) and memory usage across different backbones and anomaly detection algorithms. Reported values correspond to the best performance achieved among the evaluated strategies. The color scale encodes memory consumption (in MB) on a logarithmic scale. FT denotes fine-tuning, while JT refers to joint training. “cl 10k” and “cl 30k” indicate memory bank sizes, whereas “cl 40” and “cl 100” denote replay memory sizes. “cl multi” and “cl uni” refer to PaDiM-based variants.
Figure 3: Comparison of the considered methods across tasks in the CL setting using Mobilenet as backbone and MVTec as dataset.
Model
Image ROC
Image F1
Pixel ROC
Pixel F1
PaDiM CL (original)
0.56
0.80
0.73
0.17
PaDiM CL (ours)
0.90
0.94
0.96
0.48
Table 1: Comparison between the original PadimCL [ 6 ] and our version (here using WideResNet as backbone). Given the significant improvement in the results of our version we refer to our version as PadimCL in the rest of the paper.
Model
Strategy
MCUNet
MobileNet
PhiNet
CFA
cl 100
0.87 | 0.26 | 3.03 | 17.01
0.85 | 0.22 | 5.73 | 20.46
0.85 | 0.26 | 4.60 | 17.58
cl 40
0.86 | 0.26 | 3.03 | 8.39
0.85 | 0.21 | 5.73 | 11.84
0.85 | 0.26 | 4.60 | 8.97
PaDiM
cl multi
0.93 | 0.44 | 2.24 | 447.82
0.94 | 0.50 | 11.96 | 1469.91
0.91 | 0.40 | 6.52 | 959.62
cl uni
0.88 | 0.30 0.36 | 29.86
0.89 | 0.34 | 1.22 | 98.00
0.86 | 0.29 | 0.81 | 63.98
PaDiM Lite
cl multi
0.91 | 0.37 | 0.34 | 25.59
0.90 | 0.41 | 0.57 | 46.67
0.88 | 0.31 | 0.52 | 37.63
cl uni
0.85 | 0.15 | 0.23 | 1.71
0.84 | 0.13 | 0.46 | 3.11
0.84 | 0.15 | 0.41 | 2.51
Table 2: Benchmark results on MVTec AD across all model–backbone combinations in the continual learning setting. Results are reported as: Image F1 | Pixel F1 | GFLOPs | Memory (MB) . cl 40 and cl 100 denote replay memory sizes of 40 and 100, respectively. cl multi refers to PaDiM-CL MultiModal, while cl uni denotes PaDiM-CL UniModal. CL 10K and CL 30K correspond to PatchCoreCL configurations with total memory bank sizes of 10,000 and 30,000 patches, respectively. Finally, CL++ indicates our proposed PatchCoreCL++ method.
AD Model
Strategy
DeiT Tiny
ViT Base
Dinomaly
cl 100
0.88 | 0.45 | 3.39 | 50.87
0.89 | 0.40 | 68.56 | 575.66
cl 40
0.88 | 0.44 | 3.39 | 42.25
0.90 | 0.42 | 68.56 | 567.04
Table 3: Benchmark results on MVTec AD for the Dinomaly model in the continual learning setting. Results are reported as: Image F1 | Pixel F1 | GFLOPs | Memory (MB) .
AD Model
Strategy
MCUNet
MobileNet
PhiNet
CFA
cl 100
0.76 | 0.16 | 3.03 | 17.01
0.75 | 0.20 | 5.73 | 20.46
0.73 | 0.13 | 4.60 | 17.58
cl 40
0.76 | 0.12 | 3.03 | 0 8.39
0.75 | 0.17 | 5.73 | 11.84
0.72 | 0.10 | 4.60 | 0 8.97
PaDiM
cl multi
0.83 | 0.24 | 2.24 | 447.82
0.84 | 0.29 | 11.96 | 1469.91
0.82 | 0.24 | 6.52 | 959.62
cl uni
0.77 | 0.17 | 0.36 | 29.86
0.77 | 0.22 | 0 1.22 | 0 98.00
0.76 | 0.13 | 0.81 | 0 63.98
PaDiM Lite
cl multi
0.79 | 0.18 | 0.34 | 25.59
0.79 | 0.21 | 0.57 | 46.67
0.77 | 0.14 | 0.52 | 37.63
cl uni
0.74 | 0.07 | 0.23 | 0 1.71
0.73 | 0.04 | 0.46 | 0 3.11
0.74 | 0.04 | 0.41 | 0 2.51
Table 4: Benchmark results on ViSA across all model–backbone combinations in the continual learning setting. Results are reported as: Image F1 | Pixel F1 | GFLOPs | Memory (MB) .
AD Model
Strategy
DeiT Tiny
ViT Base
Dinomaly
cl 100
0.78 | 0.33 | 3.39 | 50.87
0.81 | 0.32 | 68.56 | 575.66
cl 40
0.77 | 0.33 | 3.39 | 42.25
0.86 | 0.38 | 68.56 | 567.04
Table 5: Benchmark results on VisA for the Dinomaly model in the continual learning setting. Results are reported as: Image F1 | Pixel F1 | GFLOPs | Memory (MB) .
Figure 4: Evaluation of VAD models across different backbones on the MVTec Dataset. The charts compare F1 Pixel Level, F1 Image Level, computational complexity (GFLOPS), and Memory (MB).
Visual Anomaly Detection (VAD) is crucial for industrial inspection, yet most existing methods are limited to single-category scenarios, failing to address the multi-class and continual learning demands of real-world environments. While Teacher-Student (TS) architectures are efficient, they remain unexplored for the Continual Setting. To bridge this gap, we propose AdapTS, a unified TS framework designed for multi-class and continual settings, optimized for edge deployment. AdapTS eliminates the need for two different architectures by utilizing a single shared frozen backbone and injecting lightweight trainable adapters into the student pathway. Training is enhanced via a segmentation-guided objective and synthetic Perlin noise, while a prototype-based task identification mechanism dynamically selects adapters at inference with 99% accuracy. Experiments on MVTec AD and VisA demonstrate that AdapTS matches the performance of existing TS methods across multi-class and continual learning scenarios, while drastically reducing memory overhead. Our lightest variant, AdapTS-S, requires only 8 MB of additional memory, 13x less than STFPM (95 MB), 48x less than RD4AD (360 MB), and 149x less than DeSTSeg (1120 MB), making it a highly scalable solution for edge deployment in complex industrial environments.
Manuel Barusco, Davide Dalle Pezze, Francesco Borsatti +1
Visual Anomaly Detection (VAD) is essential for industrial quality control, enabling automatic defect detection in manufacturing. In real production lines, VAD systems must satisfy strict real-time and privacy requirements, necessitating a shift from cloud-based processing to local edge deployment. However, processing data locally on edge devices introduces new challenges because edge hardware has limited memory and computational resources. To overcome these limitations, we propose two efficient VAD methods designed for edge deployment: PatchCore-Lite and Padim-Lite, based on the popular PatchCore and PaDiM models. PatchCore-Lite runs first a coarse search on a product-quantized memory bank, then an exact search on a decoded subset. Padim-Lite is sped up using diagonal covariance, turning Mahalanobis distance into efficient element-wise computation. We evaluate our methods on the MVTec AD and VisA benchmarks and show their suitability for edge environments. PatchCore-Lite achieves a remarkable 79% reduction in total memory footprint, while PaDiM-Lite achieves substantial efficiency gains with a 77% reduction in total memory and a 31% decrease in inference time. These results show that VAD can be effectively deployed on edge devices, enabling real-time, private, and cost-efficient industrial inspection.
Arianna Stropeni, Fabrizio Genilotti, Francesco Borsatti +3
Continual anomaly detection (CAD) addresses the need for industrial inspection systems to adapt to evolving production conditions, yet existing methods share three critical gaps: unrealistic evaluation, no systematic comparison, and no consideration of edge deployment constraints. We introduce a unified benchmark combining discrete-task evaluation on structural and logical anomalies, a novel continuous drift protocol, the first head-to-head comparison of all published CAD methods, and computational efficiency profiling on edge hardware. Our results reveal that existing CAD methods do not consistently outperform traditional approaches with simple experience replay. Thus motivated, we propose DINOSaur, a training-free method combining a frozen DINOv3 backbone with spatially-indexed coreset memory and neighborhood-restricted anomaly scoring. DINOSaur achieves zero forgetting by construction, outperforms all evaluated methods across all five protocols, and runs at sub-100,ms inference on an NVIDIA Jetson Orin Nano, with on-device adaptation to new tasks in under 30 seconds.