Smoke Video Detection Algorithm Based On 3D Convolutional Neural Network

Smoke detection can play a key role in the early stage of a fire. Traditional video smoke detection mainly extracts image features of smoke and uses machine learning methods to classify and recognize smoke. However, smoke video is an image of continuous frames, and adjacent frames of images are related to each other. If time complexity is ignored and only spatial features are considered, the key features cannot be fully extracted. By combining the 3D convolutional neural network and introducing the attention mechanism, the temporal and spatial features of the smoke can be effectively weighted and merged, and the 3×1×1 and 1×3×3 convolution kernels are used instead of the 3×3×3 convolution kernels. Reduce model parameters. In order to eliminate most of the non-smoke areas to reduce the time complexity of the system, the MobileNetV2-SSD algorithm is used to do a priori score to locate the suspected smoke areas. Experimental results show that compared to traditional smoke detection methods and 2D smoke detection convolutional networks, the accuracy of smoke detection has been improved.