A Dataset Preparation and Augmentation Method for Improving Infrared Small-Object Detection Algorithms Based on Neural Networks

Abstract

This paper introduces a deterministic, spatially constrained dataset augmentation framework explicitly designed to support convolutional neural network architectures for infrared small-target detection. The framework is tailored for targetdetection algorithms that, instead of evaluating entire image frames globally, focus on processing localised image regions sequentially or in parallel, such as sliding-window or bipyramidal approaches. While conventional stochastic augmentations-including arbitrary cropping, blurring, or geometric distortions-disrupt the physical sensor noise floor and dilute target signatures, the proposed framework enforces operational parity by aligning augmentation parameters directly to the target geometric bounds and empirical statistical properties of the dataset. The framework consists of methods for extracting multi-tier statistical distributions from a dataset, executing flexible augmentation configurations, and utilising a comprehensive telemetry log to record training hyperparameter sets and inference results, enabling precise traceability of neural network detection accuracy to multiple input parameters. Crucially, the approach imposes no structural constraints on the underlying neural network models, allowing researchers to independently vary neural network topologies and parameters while maintaining a complete historical registry of results.