Smoke detection method and system based on foreground and background analysis and electronic equipment
A technology of foreground background and detection method, applied in the field of smoke detection, can solve the problems of false negatives and false positives, and achieve the effect of reducing false negatives and false positives, and improving sensitivity and accuracy.
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Example Embodiment
[0036]Example 1
[0037]In this embodiment, the smoke detection still uses deep learning, but the adversarial generation technology is added, including three sub-networks: smoke detection network D, smoke removal network G, smoke analysis network S, and 4 difference functions, smoke motion loss LS, Background difference LB, Scene recovery difference LRAnd background legacy LDS ,Such asfigure 1 Shown. Among them, the total network loss, extracting frame I from the video stream from the camerat, Generate foreground image through smoke detection network DGenerate background image through smoke removal network GForeground imageAnd backgroundImage synthesis new scene imageForeground image of current frameGenerate smoke motion loss L through smoke analysis network SS.
[0038]Background image of current frameBackground image with previous frameThe difference isAdd up and leave L for backgroundB,which is:
[0039]
[0040]Foreground image generated by smoke detection network DAnd smoke removal network...
Example Embodiment
[0049]Example 2
[0050]This embodiment discloses a smoke detection network D, which uses a codec network to generate a foreground image from the current frame ItThe network parameter is θD, This process can be described as:
[0051]
[0052]The encoder contains four convolutional layers, which transform the image It into high-dimensional features with a size of 1 / 24 and 512 channels. The decoder uses 4 symmetrical deconvolution layers to convert from high-dimensional features to smoke channel images of the original sizeThe size of the convolution kernel of each layer is 5, the stride is 2, using leaky ReLU and batch normalization operations. The corresponding layer of the decoder and the encoder is connected with a jump layer to store advanced information to ensure that the up-sampled pixels can resolve the smoke with high quality. The network structure is likefigure 2 As shown, the smoke is gradual, and the generated smoke channel image does not have a large gradient, so the smoke detectio...
Example Embodiment
[0055]Example 3
[0056]This embodiment discloses a smoke removal network G, which uses a codec network to obtain information from the current frame ItGenerate background imageThe network parameter is θG, The process is described as
[0057]
[0058]The encoder contains 8 convolutional layers, which transform the input into high-dimensional features of 512 channels, and the decoder uses 8 symmetrical deconvolutional layers to convert the high-dimensional features into original size background imagesIf the encoder keeps reducing the resolution, and then the decoder keeps increasing the resolution, the only way to deal with this will cause the gradient to disappear and the information imbalance of each layer. In fact, the background content of the surveillance camera rarely changes. The interlayer connection allows this part of information to be directly used for generation without the need to go through the calculation process of compression and decompression to avoid the loss of information....
PUM
Abstract
Description
Claims
Application Information
- R&D Engineer
- R&D Manager
- IP Professional
- Industry Leading Data Capabilities
- Powerful AI technology
- Patent DNA Extraction
Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic.
© 2023 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap