A latency-aware multispectral image processing framework for real-time band compression and spectral-spatial optimization

Real-time multispectral image processing requires a design criterion that is different from offline spectral modeling: the selected bands, features, and classifier must be evaluated by both recognition performance and end-to-end runtime. This paper presents a latency-aware framework that converts a physical multispectral acquisition into a virtual inspection object composed of an image cube, an automatic region-of-interest mask, spectral-spatial descriptors, and measured runtime stages. The framework compares full-band and compact-band pipelines under leakage-controlled repeated cross-validation, and it includes a latency-aware compact band selection rule that balances mutual-information relevance, inter-band redundancy, and acquisition cost. The experimental case uses a 31-band multispectral image set consisting of 100 sample-level objects, divided into four groups with 25 physical samples per group. Each sample contains 31 aligned band images with a spatial resolution of 2752 by 2192 pixels. The region of interest (ROI) is automatically extracted from the reference band through a segmentation pipeline, and feature modes including spectral mean, full spectral statistics, texture, and their fusion are evaluated within the latency-aware framework. The full 31-band reference obtained 0.871 ± 0.066 macro-F1 with a measured latency of 1556.7 ms per sample. A 20-band compact pipeline retained 0.826 ± 0.075 macro-F1 while reducing latency by 30.1%, and a 10-band compact pipeline retained 0.808 ±0.057 macro-F1 while reducing latency by 64.4%. The fastest latency-aware three-band operating point reached 227.5 ms per sample with lower macro-F1. These results show that explicit accuracy-latency analysis is necessary for selecting realistic operating points in real-time multispectral inspection.