Abstract: During field acquisition, near-infrared hyperspectral imaging (NIR-HSI) systems are prone to stripe noise caused by detector response non-uniformity, which distorts the spectral curve shape and consequently degrades recognition performance. To efficiently discriminate among three visually similar forest-type camouflage nets (multispectral, hot-spot optical, and anti-scratch hot-spot optical variants), a high-frequency residual correction convolutional neural network (HRC-CNN) that balances denoising effectiveness and computational cost was proposed in this study. First, a lightweight high-frequency residual correction (HRC) algorithm was developed by exploiting the directional characteristics of stripe noise, treating it as a high-frequency component for targeted suppression and maximizing the preservation of spatial texture information in the hyperspectral images. A lightweight 3D-CNN was then constructed on the corrected hyperspectral data cube to perform joint spatial–spectral classification learning. Experiments were conducted on five field NIR hyperspectral camouflage-net datasets acquired using a 900-1700 nm NIR-HSI system. Results showed that, with an overall accuracy above 0.99, HRC-CNN achieved a mean mIoU of 0.5602, representing a 12.26% improvement over the baseline 3D-CNN. These findings indicate that the proposed method effectively alleviates stripe-induced false alarms and missed detections, enables the reliable recognition of three classes of visually similar forest-type camouflage nets in complex natural backgrounds, and provides a useful reference for NIR hyperspectral camouflage net recognition.