A fog-aware lightweight vision transformer architecture for real-time coastal landscape recognition on edge devices

Engineering notes

Engineering notes will be added by the Power for Solar editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为光伏效率、钙钛矿太阳能电池、储能技术、太阳能热利用、BIPV、并网技术等高价值论文补充中文说明。

Original abstract

Intelligent landscape recognition sits at the heart of smart coastal tourism, yet pushing vision Transformers onto resource-constrained edge hardware remains awkward-the self-attention mechanism scales quadratically with token count, and marine atmospheric haze quietly erodes recognition accuracy long before it is noticed in the laboratory. We propose a hybrid lightweight architecture that swaps standard self-attention in the early stages for depthwise separable convolution token mixers, switches to kernel-based linear attention in the deeper stages, and follows an adaptive channel reduction policy that trims feature dimensions where redundancy is empirically highest. A Fog-Aware Feature Calibration module, motivated by the Koschmieder atmospheric scattering model, is embedded between the convolutional and attention stages so that learned dehazing happens inside the network rather than as a separate preprocessing pass. Training proceeds through a three-phase pipeline that interleaves progressive structured pruning with temperature-scheduled knowledge distillation from a Swin-Base teacher; the composite objective shrinks the model to 4.8 M parameters and 0.91 GFLOPs without abrupt accuracy collapse. On a newly collected 17,565-image dataset spanning eight coastal scene categories across six shoreline regions of Zhejiang Province, our model reaches 92.6% ± 0.4% Top-1 accuracy (mean ± std over three independent runs), the best result among all baselines below 6 M parameters, including the recently released MobileViT-v2, FastViT, TinyViT, EfficientViT and RepViT. On an NVIDIA Jetson Orin Nano with TensorRT INT8 optimisation the system delivers 18.3 ms latency (54.6 FPS) within a 7.4 W average power envelope, and a 200 m visibility heavy-fog subset that mixes synthetic and real captures shows an 85.7% accuracy retention rate-6.8 percentage points above the strongest baseline. Cross-dataset zero-shot tests on Fujian, Hainan and the Places365 coastal subset, together with deployment benchmarks on Jetson Nano 4 GB and Coral Edge TPU, suggest the design transfers reasonably-though not effortlessly-beyond the original collection sites.

Links and sources

• PDF from original source

Comments