Objective The accurate and real-time segmentation of camouflaged personnel in complex dynamic environments is a critical yet challenging task for modern detection and perception systems. Multispectral filter array (MSFA) cameras have emerged as a promising solution due to their ability to capture spectral information in a single exposure, offering a lightweight alternative to traditional mechanical multispectral systems. However, existing segmentation pipelines for MSFA images typically rely on a two-stage architecture that first performs demosaicing to reconstruct full spectral information, followed by segmentation using a deep neural network. This demosaicing step, whether based on interpolation or deep learning, introduces spectral distortion, edge blurring, and computational overhead, leading to irreversible error propagation and increased latency. Such limitations make conventional approaches unsuitable for highly dynamic scenarios where both high precision and real-time processing are essential. To address these challenges, this paper proposes a novel end-to-end single-stage segmentation framework that directly processes raw MSFA mosaic images, eliminating the need for demosaicing and achieving superior performance in both accuracy and efficiency.

Methods The proposed method, termed MSL-YOLO (Multispectral YOLO), is a single-stage network designed for direct camouflaged target segmentation from raw MSFA images. The network architecture consists of three main components: backbone, neck, and head. A key innovation is the Extract Module, which leverages the physical layout of the 3×3 periodic MSFA filter array to decompose the input single-channel mosaic image into nine distinct single-band feature maps, each corresponding to a specific spectral channel. This operation preserves the original spectral signatures without interpolation or reconstruction. To enhance spatial-spectral feature representation, the backbone incorporates a lightweight attention mechanism, C3CSAM (C3-Channel Spatial Attention Module), which integrates channel-wise and spatial-wise attention within grouped feature subspaces, followed by a weight-refinement strategy to emphasize discriminative responses. The neck employs a Feature Pyramid Network (FPN) combined with a Path Aggregation Network (PAN) to enable robust multi-scale feature fusion, while the head performs pixel-level segmentation by fusing mask coefficients with prototype masks. The model is trained from scratch on a newly constructed dataset, MSCD (Multispectral Camouflage Personnel Dataset), which comprises 557 raw MSFA images captured in diverse outdoor environments, including variations in terrain, lighting, distance, pose, and occlusion.

Results and Discussions Experimental results demonstrate that MSL-YOLO achieves state-of-the-art performance on the MSCD dataset, with a mean average precision (mAP) of 98.7% and an inference speed of 148 frame/s, corresponding to a latency of only 6.8 ms per image. In contrast, traditional two-stage approaches combining demosaicing with YOLOv5 segmentation yield significantly lower performance. For instance, the method using bilinear interpolation demosaicing achieves only 96.1% mAP at 25.5 frame/s (39.2 ms), while the method using a deep learning-based demosaicing network (MCAN) achieves 97.9% mAP but at a lower speed of 20.4 frame/s (48.9 ms). Even when MSL-YOLO is used as the segmentation model in a two-stage setup, the mAP reaches 98.1% but with a latency of 56.3 ms, highlighting the bottleneck introduced by demosaicing. Qualitative results further illustrate the robustness of the single-stage approach. Two-stage methods often produce false positives in cluttered backgrounds, miss partially occluded targets, and generate fragmented boundaries under challenging lighting conditions. In contrast, MSL-YOLO consistently delivers cleaner, more accurate segmentations with well-defined target contours. Ablation studies on attention mechanisms show that the proposed C3CSAM module achieves the best trade-off between accuracy and efficiency (98.7% mAP, 3.2 GFLOPs, 148 frame/s), outperforming other attention modules such as SE, CBAM, CA, NAM, and GAM. Additional experiments on group partitioning in C3CSAM reveal that setting the number of groups to four yields optimal performance, while cross-validation results confirm the model’s robustness under limited sample conditions, with a mean accuracy of 87.3% and a standard deviation of 2.04%.

Conclusions This paper presents a single-stage segmentation framework, MSL-YOLO, designed specifically for raw multispectral mosaic images captured by MSFA cameras. By eliminating the conventional demosaicing step, the proposed method avoids reconstruction errors, preserves original spectral information, and significantly reduces computational overhead. The integration of the Extract Module and the lightweight C3CSAM attention mechanism enables effective joint learning of spatial and spectral features directly from the mosaic input. Comprehensive experiments on the newly constructed MSCD dataset demonstrate that MSL-YOLO achieves a superior combination of high segmentation accuracy and real-time processing speed, outperforming traditional two-stage pipelines in both metrics. This work provides a practical and efficient solution for real-time visual target segmentation in complex dynamic environments, with strong potential for deployment in resource-constrained platforms. Future work will focus on expanding the dataset to include more samples and diverse scenarios to further enhance model generalization.