Two-stage smoke recognition convolutional neural network combining color and texture features

Abstract

The invention discloses a two-stage smoke recognition convolutional neural network combining color and texture features. The network is mainly composed of a color channel convolutional sub-network anda texture convolutional sub-network. Operating the color channel convolution sub-network in a color space, and extracting a maximum difference color mode between smog and non-smog types; the textureconvolution sub-network operates in a texture space, and shape features are extracted on a color channel. The last layer of the whole network comprises features pooled from the upper layer, features pooled from the middle layers such as a color layer and a texture layer are spliced, then the features on each channel are globally maximum pooled into a one-dimensional vector, and whether smoke and flames exist in a scene is judged through sigmoid function classification. The color channel generated by automatic training can cover more samples; compared with a traditional manual feature extraction method, the method is higher in universality, lighter than a general convolutional network and higher in accuracy.

Description

technical field [0001] The invention relates to the technical field of smoke recognition, in particular to a two-stage smoke recognition convolutional neural network combining color and texture features. Background technique [0002] Fire has caused great damage to human production and life. The later the fire warning time, the greater the casualties and the higher the property damage. As smoke is an important feature of the initial fire, if the smoke can be effectively captured through visual devices, it can provide timely and effective early warning before the fire has expanded, thereby reducing casualties and property losses. Since surveillance cameras are easy to set up and exist widely, it is of great practical significance to carry out vision-based smoke detection / fire research. [0003] Early smoke recognition mostly revolved around static features such as color, texture, and outline. For example, smoke recognition models were established by analyzing color informat...

AI Technical Summary

Problems solved by technology

However, using a general-purpose convolutional neural network, the network will be too large, and often the samples are not sufficient, and the strengths of the deep network cannot be fully utilized.

Wang Zhenglai, Oleksii Maksymi, and Yi Zhao are aware of the role of color, texture, and outline, and consciously combine these information into the network, but these networks are often divided into two parts, the first part adopts traditional methods, and the second part uses Convolutional neural network is used to judge, while the workload is large, the accuracy is low

Images