Pyramidal inhibitory lightweight autoassociative reconstruction for cost-constrained visual anomaly detection

Visual anomaly detection plays an important role in industrial inspection, edge monitoring and visual quality control, particularly where abnormal samples are rare, heterogeneous, and often unavailable during training. Recent deep anomaly detection approaches have achieved high classification performance by leveraging pretrained backbones, memory banks, nearest-neighbor search, synthetic anomaly generation, or auxiliary discriminators. However, these methods often increase parameter count, floating-point operations (FLOPs), memory usage, update size, and energy consumption, which complicates deployment on cost-constrained industrial or edge hardware. This study introduces PILAR (Pyramidal Inhibitory Lightweight Autoassociative Reconstruction), a compact reconstruction-based architecture for anomaly detection. PILAR employs a pyramidal autoassociative encoder-decoder with local receptive fields, a spatial bottleneck, pure-pyramid reconstruction, and lateral inhibition to promote local competition within feature maps. The model is trained only with normal samples, and the anomaly score is obtained from reconstruction residuals. Thus, during inference, PILAR does not require class prototypes, covariance matrices, support vectors, density parameters, nearest-neighbor databases, or discriminator states. PILAR is evaluated on MNIST, Fashion-MNIST, CIFAR-10, MVTec AD, and VisA datasets using direct image reconstruction, whole-image RGB reconstruction, patchwise RGB reconstruction, and Feature-PILAR-FCN, a fully convolutional feature-space variant that uses compact PILAR heads to reconstruct selected frozen feature maps. In the evaluated implementation, these maps are taken from ResNet18 layer 2 and layer 3 prior to residual scoring. The proposed detector reaches 98.86± 0.04% Image-AP and 92.63± 0.05% AUPRO@30 on MVTec AD over five seeds; the unchanged industrial configuration reaches 92.53± 0.07% Image-AUROC on VisA. Compared with the evaluated ResNet18 configurations of SimpleNet and GLASS, Feature-PILAR-FCN retains 35.6% fewer inference parameters and achieves 5.2–5.6 times their throughput at batch size 16 on the same Apple M4 Max GPU in FP32, despite higher single-image latency. These results indicate that PILAR can be used as a low-state anomaly detector when memory footprint and computational cost are relevant constraints. They also show that direct RGB reconstruction is useful mainly in selected low-cost screening cases, while broader industrial detection requires a more informative representation, which in this work is obtained from frozen feature maps.