CPAD: Continuous Pretraining for Infrared Images With Advances in Data, Preprocessing, and Paradigm

Infrared remote sensing imagery has emerged as a critical data source in environmental perception and intelligent monitoring, with significant potential in scenarios requiring robust performance under low-light, nighttime, or complex atmospheric conditions. Improving the perception capability of vision models for infrared images is essential to these applications. However, existing vision models pre-trained on RGB data often underperform in the infrared domain due to the severe distribution shift. To bridge this gap, we systematically investigate self-supervised learning for infrared imagery from three perspectives: data, preprocessing, and pre-training paradigm. This work identifies three core challenges hindering effective infrared pre-training: 1) the lack of large-scale infrared pre-training datasets, 2) the non-iconic characteristic of infrared images, and 3) the scarcity of fine-grained textures. To mitigate the data scarcity, we construct a large and diverse dataset called Multi-Scene Infrared Pre-training (MSIP), consisting of 178,756 images. As for the non-iconic characteristic, a simple yet effective image preprocessing method, Object-Sensitive Random RoI Cropping (OS-RRC), is introduced. To alleviate the negative impact of weak textures on feature learning, we propose a novel pre-training paradigm called Continual Pre-training with ADapter (CPAD), which incorporates adapters to learn domain-specific features while freezing off-the-shelf pre-trained models to retain the general feature extraction capability. Furthermore, to overcome limitations of vanilla adapters in CPAD, we introduce the patchwise-scale adapter equipped with dynamically learnable scale factors, allowing for pre-training once and subsequent application to various tasks. Extensive experiments show that the proposed framework achieves notable performance gains across three representative downstream tasks, validating its effectiveness for infrared image analysis in environmental monitoring and intelligent perception systems.