Research on anomaly detection models for optoelectronic sensing images based on deep feature fusion

Artificial intelligence-driven deep feature fusion methods hold significant value in anomaly detection for optoelectronic sensing images. This paper constructs a deep feature fusion model based on multi-scale convolutions, spatial-channel attention, and local-global dual-domain modeling. Building upon the ResNet50 (2015) backbone architecture, it introduces parallel scale paths and a Transformer encoder to achieve cross-scale coupled representation of high-dimensional semantics. The model enhances features through 3×3/5×5/7×7 convolutions, multi-level SE-CBAM attention, and weighted fusion strategies, forming a highly robust recognition framework. Experiments on the MVTec-AD and MTAD datasets demonstrate significant improvements over the baseline CNN model: accuracy increases from 0.882 to 0.932, F1-score rises from 0.854 to 0.922, and AUC climbs from 0.887 to 0.961. achieving an accuracy of 0.946 and an F1-score of 0.943 on the multimodal MTAD dataset, validating the fusion strategy's significant advantage in cross-scenario recognition.