A Smoke Detection Method Based on Multi-Network Model Fusion

A technology of model fusion and detection methods, applied in character and pattern recognition, instruments, calculations, etc., can solve problems such as difficult to distinguish well, insufficient discrimination ability, loss of original images, etc., to improve generalization ability and avoid under-fitting problems and reduce the effect of false alarms

CN109948557BActive Publication Date: 2022-04-22NAT UNIV OF DEFENSE TECH
5 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Publication Date
2022-04-22

Smart Images

  • Figure 1
    Figure 1
  • Figure 2
    Figure 2
  • Figure 3
    Figure 3
Patent Text Reader

Abstract

The invention relates to a smoke detection method for multi-network model fusion. Fusion of VGG16 and ResNet50 network models to achieve reliable detection of smoke. The fusion network proposed by the present invention can extract richer smoke image detail features, and enhance the feature's ability to distinguish between smoke images and smoke-like images. Using the feature transfer learning method based on the homogeneous space, the feature extraction layer of the pre-trained VGG16 and ResNet50 models can be well transferred to the classification task of the target data set in the smoke scene, and the generalization ability of the model can be improved at the same time. Through the fusion of multi-network models, the ability to distinguish features is enhanced, the false alarm phenomenon caused by objects similar to smoke such as clouds and fog is reduced, and the reliability of smoke detection is further improved.
Need to check novelty before this filing date? Find Prior Art

Description

technical field

[0001] The invention relates to a smoke detection method for multi-network model fusion. Background technique

[0002] Fire early warning technology based on computer vision is playing an increasingly important role in the field of fire detection and early warning. Among them, smoke detection is of great significance to the early detection of fire. At present, smoke detection methods are mainly divided into two categories. One is to realize fire detection based on traditional features such as color, shape, texture, and motion. For example, the document "Smoke detection invideo using wavelets and support vector machines" uses discrete wavelet transform to extract smoke image features , and apply a support vector machine for classification. The document "Smoke Detection in Video Sequences: Combined Approach" uses the method of area matching to extract the suspected smoke area, and uses the color change speed of the background object to identify the smoke. Th...

Examples

specific Embodiment

[0022] Step1: Use the camera to collect the scene monitoring video;

[0023] Step2: For each frame of image in the video, use the bilinear interpolation method to scale the image size to 224×224;

[0024] Step3: Use the VGG16 feature extractor and the ResNet50 feature extractor to extract features, and place the 7×7×2048=100352-dimensional features extracted by the ResNet50 network after the 7×7×512=25088-dimensional features extracted by the VGG16 network, and construct a 100352+25088=125440-dimensional features;

[0025] Step4: Treat each feature as a neuron node, connect the extracted features in a fully connected (FC) manner, and output 1024 neuron nodes;

[0026] Step5: Use the Dropout method to randomly select neurons according to a certain probability p (where p=0.3) and discard them;

[0027] Step5: The remaining neurons are still connected in a fully connected (FC) manner, outputting 128 neuron nodes;

[0028] Step6: Still using the Dropout method, randomly select ...