Research On Fire Image Detection Method Based On Transfer Learning

Fire is one of the disasters that seriously threaten the safety of our lives and property and social development.The traditional fire alarm system is affected by many factors,such as high false alarm rate and untimely alarm.How to improve these problems and make the fire alarm system more practical has become a hot research topic.The traditional fire detection technology based on image processing is to identify the flame by extracting the characteristics of flame or smoke,and the current feature extraction algorithm has the problems of high false detection rate and low real-time performance,the current fire detection technology based on deep learning is trained through the data set A discriminant model for discriminating flames or smoke is developed to achieve the purpose of flame recognition,but there are also problems such as few data sets,slow prediction speed,and low model accuracy.In response to these problems,this paper proposes a fire image detection method based on transfer learning.1.This paper first analyzes several commonly used moving target detection methods,and selects the Vi Be algorithm to detect flame motion prospects through experiments.This method can quickly establish a model to achieve the segmentation of flame prospects;then analyzes several commonly used color models,Improved the color model discrimination decision based on the combination of RGB and HSI,the improved model can extract the flame area more completely,and the false detection rate is low;finally,the fire area discrimination based on the improved Vi Be and color double threshold is proposed.Introduce gridding in the image,combine the movement characteristics of the flame with the color model in the grid,and mark the area determined as a fire.By calculating the connectivity domain,the position of the early flame can be accurately located.2.Use the deep transfer learning method to transfer the trained model in the source domain to the field of fire detection.First obtain the original video frame image as a data set,and perform preprocessing for cropping,data enhancement,and normalization;Then train 5 common network models such as Inception V3,Resnet18,Resnet50,Dense Net121 and Densenet169 pre-trained on the Image Net dataset,the training data set is a pre-processed frame image set;then for model testing and evaluation,the model with the best migration ability is selected for fire detection.Finally,the selected optimal model is compared with the existing Faster-RCNN and Inception V1-Onfire models.The overall classification effect is better,and it meets the real-time requirements.In summary,the fire image detection method based on transfer learning proposed in this paper can quickly and effectively detect fires and issue more accurate fire point location alarms.