Research On DC-ILSTM Network Forest Smoke Identification

Forests are one of the most important resources on earth.Not only can it provide various valuable resources for life and production,but more importantly,it can protect the environment and maintain the balance of the ecosystem.However,the occurrence of forest fires has caused serious damage to the forest,such as burning a large number of trees,damaging forest life and destroying forest replacement,resulting in the imbalance of the forest ecosystem and even threatening the safety of people’s lives and property.Therefore,rapid identification of smoke is an important means to prevent the occurrence of hazards.This paper uses deep learning methods to study forest fire smoke recognition.In view of the great similarity of the smoke characteristics of each frame sampled,as well as the relatively small forest fire smoke data set and the single scene,in order to fully utilize the static and dynamic information of smoke to achieve the purpose of preventing forest fires,this article separately research on learning and long-term and short-term memory networks,the specific contents are as follows:1.Propose a feature transfer deep convolution(FT-DC)smoke detection model.The smoke data set is collected and generated,and a smoke detection model based on FT-DC is proposed.The fire smoke recognition method for computer vision is easily interfered by the external environment,the recognition scene is single,and the forest fire smoke data set is relatively small.This paper combines deep learning and transfer learning,and uses the feature isomorphic transfer learning method to extract smoke spatial characteristics.First preprocess the smoke image,divide all the images into categories of uniform size,and expand the data set through random rotation,shearing and flipping operations;secondly pre-train the VGG-16 model to obtain the convolutional layer that has been trained on the Image Net data set parameters;then connect the softmax activation layer fine-tuning model,use the smoke data set as input for model fine-tuning training;finally use the public data set and the extended data set for experimental verification.Experimental results show that the feature recognition method can achieve a smoke recognition rate of more than 93.3%,but the recognition rate is only about 50% in slow-moving smoke videos and smoke-like environments,so 2D convolutional neural networks are not suitable for video Smoke detection.2.Propose a fire smoke recognition model algorithm based on Deep Convolution Integrated Long Short-Term Memory(DC-ILSTM)network.An integrated Long Short-Term Memory(ILSTM)network was developed,and a fire smoke recognition model based on DC-ILSTM network was designed.In view of the problem that the two-dimensional convolutional neural network can only extract smoke spatial information from video images,resulting in false positives and false negatives,this paper uses a recurrent neural network to learn the smoke spatial behavior,and proposes the ILSTM network to aggregate the static characteristics of smoke to perform two Secondary recognition.First,perform feature preprocessing,segment the extracted smoke feature sequence,map it to the range of [0,1],and obtain the maximum feature value in the region through the maximum pooling layer;then complete LSTM serialization learning,The processed smoke feature sequence is trained as the input of the LSTM network to complete the expression of the feature in the time series;finally,the softmax layer is connected for binary classification and the smoke recognition result is output.The experimental results show that DC-ILSTM significantly improves the recognition effect.The model detects smoke with a smaller number of earliest frames,and the detection accuracy is above 94.5%,and the false alarm rate is as low as 1.30%.The DC-ILSTM algorithm proposed in this paper is experimented on the smoke data set.Compared with other deep convolutional long recurrent neural networks,the experimental results show the effectiveness of feature transfer and the combination of smoke dynamic information to improve the accuracy of smoke recognition and have a good applicability in forest fire smoke detection.