多光譜融合的紅外艦船目標(biāo)輕量化檢測(cè)

打開文本圖片集

關(guān)鍵詞：紅外艦船；多光譜目標(biāo)檢測(cè)；輕量化；模型剪枝；知識(shí)蒸餾中圖分類號(hào)：TN291；TP391 文獻(xiàn)標(biāo)識(shí)碼：Adoi：10.37188/OPE.20253308.1327 CSTR：32169.14.OPE.20253308.1327

Abstract： In order to solve the problems of large size，low eficiency，and high deployment requirements for embedded devices of infrared multi-spectral ship targe detection models，a lightweight ship targe infrared multi-spectral detection model YOLOv8n-MFLW was proposed. Firstly，the model replaced the backbone network with a lightweight network，HGNetv2. Based on GSConv convolution，the modules of HGBlock and C2f were reconstructed to reduce the model parameter count while retaining the model's feature extraction and fusion capabilities. A self-adaptive sparse structured pruning algorithm，La-Depgraph，was proposed to prune the model，leading to a significant reduction in the model's parameters. Finally，an intermediate feature knowledge distilation learning strategy was employed to recover the accuracy loss caused by pruning and enhance the model's detection performance.Experimental results show that compared to the original model，the improved lightweight ship targe infrared multi-spectral fusion detection model achieves a detection accuracy of 96.4% ，an increase of 1.2% . The model's parameter count，computational complexity，and memory usage are only 0.9MB ，3.5 GFlops，and 2.3MB ，respectively，re duced by 88.1% ， 81.2% ，and 82.8% . Therefore，the proposed model is small in size and high in accuracy，it has a better detection performance and is capable for ship target detection tasks in complex environments.

Key words： infrared ships；multispectral object detection；lightweight；model Pruning；knowledge distillation

1引言

在現(xiàn)代化海戰(zhàn)中，紅外成像制導(dǎo)技術(shù)因其具備全天候工作能力、強(qiáng)大的穿透力、遠(yuǎn)距離作用、反隱身、偽裝識(shí)別和抗干擾能力等多項(xiàng)優(yōu)點(diǎn)而得到廣泛應(yīng)用。（剩余13059字）