Fire detecion and smoke detection

asked 2018-10-16 00:24:53 -0600

aayushkt gravatar image

updated 2020-12-02 11:32:50 -0600

I want to create 2 binary classifier model

1) Fire detection

2) Smoke detection

Can you please suggest which will be better to train LBP cascade(It require annotated data) or SVM/Logistic regression/ or any other deep learning binary classifier(Inception-v3, Dense-net)?

I have a positive and negative dataset of both fire and smoke dataset.

My concern here is speed(accuracy around 80% is acceptable)!

One of the test image is as bellow.

Test image

edit retag flag offensive close merge delete

Comments

1

Well the task you want to solve is classification. Forget about Regression. Classification is : "Thats a fire, Thats smoke" Regression is "There will be fire in the next n hours"

Binary Classification is "That fire or nor" , "That smoke or not". To train a classifier you need labeled data.

holger gravatar imageholger ( 2018-10-16 00:33:06 -0600 )edit

throwing "buzzwords" around won't help.

berak gravatar imageberak ( 2018-10-16 01:56:22 -0600 )edit

I hope we both are on same page, logistic regression is also a binary classifier.And yes i can go with binary classification, just wanted to know which direction is worth investing time whether LBP cascade or any other binary classifier.

aayushkt gravatar imageaayushkt ( 2018-10-16 02:18:49 -0600 )edit
1

does it have a reproducable (grayscale) texture ? does it have a "shape" ? or "edges" ? do you know, how LBP features work ?

berak gravatar imageberak ( 2018-10-16 02:31:57 -0600 )edit
1

besides classifiers, you'll have to think about features to classify

berak gravatar imageberak ( 2018-10-16 02:32:56 -0600 )edit

I do not have any idea about working of LBP. Can you please suggest will in work on fire images? Yes i will think about using features.

aayushkt gravatar imageaayushkt ( 2018-10-16 03:49:54 -0600 )edit

why don't you just hoogle it ?

berak gravatar imageberak ( 2018-10-16 03:56:51 -0600 )edit

Sure will hoogle it :p. Thanks for your response.

aayushkt gravatar imageaayushkt ( 2018-10-16 04:24:32 -0600 )edit

Old answer, but related: Smoke detection.

There are some video results.

Eduardo gravatar imageEduardo ( 2018-10-16 04:58:46 -0600 )edit

Some general thoughts: A fire varies from small and no smoke (gas stove) to big and a lot of smoke(bushfire). There are also fires with no flames but a lof of smoke(smoldering fire).

From my point of view the features (berak adviced us to think about them) of a fire are flames and smoke. They usually occur togther, but having one of them is already an indicator and worth an alert. I would train a model which would contain all these combinations of flames and smoke and label to a class "fire". I personally would let the model decide when a fire is present or not.

Your approach is to make classifiers for the features and make the decision on your own(fire or not - you are the classifier more or less). Thats could work but maybe you are missing features(less accuracy) with approach.

holger gravatar imageholger ( 2018-10-16 08:09:06 -0600 )edit