How Can an Ultra-Light Algorithm Recognize Satellite Components?

In the satellite world, knowing what’s inside is crucial for maintenance and observation missions. The new M4MLF‑YOLO model, built on YOLOv5, is designed to run on the limited resources of spacecraft. Its secret? A MobileNetV4 backbone that minimizes calculations, and an MFAC module that analyzes images in the frequency domain, sharpening component edges. Moreover, Linear Deformable Convolution (LDConv) simply adjusts the filter’s receptive field, making feature extraction more efficient. The result is a model that segments solar panels, antennas, thrusters and other parts with 95.1 % precision and 88.3 % recall, slightly beating previous versions. Parameter count dropped by 36.5 % and FLOPs by 24.6 %, making it ideal for onboard use. The team also built a mixed dataset of real and synthetic photos with varied backgrounds and lighting to test robustness. In practice, M4MLF‑YOLO shows high accuracy can coexist with lightweight design, paving the way for more autonomous robotics and diagnostics in space.

M4MLF‑YOLO merges several architectural innovations to balance accuracy and computational load. MobileNetV4, selected as the backbone, uses depthwise separable convolutions and a uniform scaling scheme, drastically cutting operations compared to ResNet or heavier YOLO versions. The MFAC (Multi‑Scale Fourier Adaptive Calibration) module transforms activation maps into the frequency domain via FFT, capturing structural patterns at multiple scales without deepening the network. This is especially effective for delineating component edges under uneven lighting, as high frequencies highlight borders. LDConv (Linear Deformable Convolution) introduces a linear parametric offset for each kernel, directly controlling receptive field size. Unlike traditional deformable convolutions with non‑linear offsets, LDConv retains computational simplicity, lowering back‑propagation costs. The C3‑Faster block, derived from CSPNet, enhances channel fusion through split‑and‑merge, limiting informational redundancy. Regarding data, the hybrid dataset includes 12 000 real LEO mission images and 8 000 synthetic ones generated with physical illumination simulations (sun‑glint, deep shadows). Images undergo histogram equalization and sensor‑specific noise augmentations, ensuring robustness to contrast and noise variations. Training uses a combined Dice‑BCE loss, optimized with AdamW (lr = 1e‑3). Results show a mAP@0.5 of 93.4 %, with a 36.5 % parameter reduction (7.2 M to 4.6 M) and a 24.6 % FLOPs drop, confirming the lightweight design’s suitability for on‑orbit applications.