Deep Learning-based Ship Fire Detection Technology Research

Fire is one of the main threats to the safety of ships.To ensure the safety of ship operations,timely and effective technical support for ship fire detection is required.Existing fire detection technology,when used in large spaces on ships,generally suffers from long response times,inefficient detection,susceptibility to interference from environmental factors and high maintenance workload,and cannot fully meet the ship’s needs for fire detection.Relying on the ship video monitoring system,the thesis introduces deep learning technology,and significantly improves the accuracy of ship fire detection by designing multi-model fusion algorithms.To address the problem of insufficient computational resources in the deployment of deep learning models,a model compression strategy combining sensitivity clipping and quantization perception is proposed,which effectively reduces the number of model parameters.The specific work is as follows.1.Analyzing the development process and main characteristics of ship fires,it is pointed out that the best time to fight a ship fire is during the heat absorption,pyrolysis and smoke generation stages,but it is also found that this window phase is very short and difficult to grasp.A comparative study has revealed that the existing fire detection techniques have certain shortcomings and need to be improved.This study further improves the video fire detection techniques which are more effective.2.Building the fire dataset.The small size of the existing fire dataset can not support the training work of deep learning sufficiently,so it needs to be reconstructed.Existing fire image data was first collected,and then a crawler program was written to crawl fire image data from major image video websites.The dataset was expanded by data augmentation,and finally the dataset was manually sorted to remove corrupted samples and to label positive and negative samples accordingly.3.By studying the back-end algorithms of existing video fire detection techniques,some problems were found: insufficient feature extraction capability,unstable detection efficiency and so on.Therefore,this study introduces deep learning techniques to do the processing of fire video images.And a deep network fire detection algorithm with multi-model fusion was designed based on the idea of integrated learning.The algorithm model designed in this study was trained and tested on a self-built fire dataset,and the efficiency of this algorithm model was verified by comparing it with traditional algorithms.4.A model compression algorithm combining sensitivity clipping and perceptual quantization is proposed to address the problems of limited storage environment and insufficient computational resources that may be encountered in the actual deployment of this algorithm model.By reducing the model network structure and changing the parameter type,the number of parameters of the model is significantly reduced and the computational complexity is decreased,while the prediction accuracy of the model does not slip significantly.In a comparison experiment with a single compression method,the compression algorithm designed in this study showed greater compression and less degradation in accuracy,showing greater practicality.The thesis constructs a large-scale ship fire dataset,designs a multi-model fusion fire detection algorithm based on deep networks,and completes model training on the self-built dataset,which significantly improves the efficiency of ship fire detection.By proposing a model compression algorithm combining sensitivity clipping and quantitative sensing,the difficulty of deploying fire detection models on the ship side is further reduced.A more efficient technical path for fire detection on ships.