基于语义分割的船舶机舱初期火灾识别算法  

Research on the Early Fire Identification Algorithm of Ship’s Engine Room Based on Semantic Segmentation

作　　者：易冠霖 吴浩峻[1] 吴韵哲 王浩亮[1] 王丹 孙定翔 YI Guan-lin;WU Hao-jun;WU Yun-zhe;WANG Hao-liang;WANG Dan;SUN Ding-xiang(Marine Engineering College,Dalian Maritime University,Dalian Liaoning 116026,China;Information Science and Technology College,Dalian Maritime University,Dalian Liaoning 116026,China;Marine Electrical Engineering College,Dalian Maritime University,Dalian Liaoning 116026,China)

机构地区：[1]大连海事大学轮机工程学院,辽宁大连116026 [2]大连海事大学信息科学技术学院,辽宁大连116026 [3]大连海事大学船舶电气工程学院,辽宁大连116026

出　　处：《船海工程》2023年第5期109-114,共6页Ship & Ocean Engineering

基　　金：辽宁省教育厅高等学校基本科研项目(LJKZ0044);大连市科技局高层次人才创新项目支持计划(2020RQ013);中央高校基本科研业务费专项资金(3132020197)。

摘　　要：为了解决经典语义分割模型对船舶机舱初期火灾中小火焰及稀薄烟雾检测能力有限的问题,提出一种改进DeepLabv3+模型算法,通过重新设计DeepLabv3+模型的解码网络,采用上采样倍数为2的近端插值逐步上采样结构,在编码网络采用更大的卷积核以增大模型感受野,并在损失函数上,提出一种新的幂次方损失函数,实验结果表明,改进后的DeepLabv3+模型与原始模型相比,mIoU总体提升了6.26%,并且可进一步提升模型对火灾初期中小火焰及稀薄烟雾的识别性能。In order to solve the problem that the classical semantic segmentation model has limited ability to detect small flames and thin smoke in the initial fire of ship cabin,an improved DeepLabv3+model algorithm was proposed,by redesigning the decoding network of DeepLabv3+model,using a proximal interpolation stepwise up-sampling structure with an up-sampling multiplier of 2,adopting a larger convolution kernel in the coding network to increase the model perceptual field,and proposing a new power-square loss function.A new power-square loss function was proposed for the loss function.The experimental results showed that the improved DeepLabv3+model improves the overall mIoU by 6.26%compared with the original model,and further improves the recognition performance of the model for small and medium flames and thin smoke in the early stage of fire.

关 键 词：船舶机舱火灾识别 语义分割 深度学习 损失函数 

分 类 号：U663.82[交通运输工程—船舶及航道工程] TP391.4[交通运输工程—船舶与海洋工程]