Train fire detection using opencv SVM

Question

I use SVM since I need ML to train my classifier, and I saw in several fire detection documents that they used SVM and logistic regression, but since there is no logistic regression in 2.4.9, I plan to use SVM. I am using opencv 2.4.9 since people say opencv 3 is buggy.

Im new to this, so it will be useful if we start with the base

I prepared several videos of fire and without fire, ready to be extracted into frames. I am new to opencv and all about classifiers. My question is, what are the basics in training the classifier, in particular, SVM, in what format do I need my images and how to train them? Are there any good links for a tutorial? I found one in the opencv documentation, but it does not teach image-based learning. What do I need to determine the parameters and for which parameters? thanks in advance

Answer 1

This is a conceptual issue that requires many articles and textbooks to explain. However, as explained in the comments, I am trying to elaborate on extracting attributes. Functional descriptors must be robust against scaling, translation, and rotation. This reliability is literally called invariant functions. For example, moments and their derivatives are one of the most famous types of invariants against rotation, scaling, and translation. You can find the use of Hu moments as described in this document . Detecting a flame or fire is something else. The function corresponds to a flame that can be extracted from the dynamic texture of the fire. For example, a fire has a special color texture that makes it stand out from the background. Conventional flame sensors use infrared sensors to detect flame. In image processing or in the RGB world, we can do the same by considering the nature of the flame itself. Flames radiate a significant portion of their energy through heat and infrared rays. Thus, it can be expected that most of the red channel will be dedicated to the flame. See the following image, for example:

In the processed image, the red channel is converted to the BW image by entering a threshold value. To be more clear, I have divided 3 channels as shown below.
R: G: B:
Obviously, the red channel says more about the flame. Therefore, we can conclude that the flame is where the R-channel has a part of its information, and then G and, finally, the B-channel. See this .
Thus, your vector function will be a three-dimensional vector, for example, for a flame path in three RGB channels. The SVM classifiers will now be ready for use. Sometimes a video may contain segments similar to flames that should be avoided, and otherwise they will lead to false alarms. SVM helps you accept or reject a candidate from a flame. To train your vector support machine, collect some true flames and some images that may be underestimated by your extractor. Then call them positive and negative traits. Finally, let opencv do the magic and train it. For more information on SVM, please watch this video Patrick Winston, Massachusetts Institute of Technology, youtube .

UPDATE ---- Since you are interested in creating vector objects, I gave you the following example. Suppose that the channels R, G, B are precisely divided, so that they can be called statistically independent, as the following; This is not true in real images in which the R, G, B planes are not statistically independent.
Therefore, a point in the RGB image will have 3 representations in the RGB channels. For example, the flame will make 3 spots on all planes R, G, B. Here, for example, the area of ​​each spot is traced. Name the flame spot in the RGB image as "A".
Representations of region A were depicted above in images R, G, B. A_r, A_g, A_b denote the corresponding region of region A on the planes R, G, B, respectively.

Therefore, point A will be represented by a triplet (Ar, Ag, Ab) in the xyz plane. SVM now accepts this vector as input and decides if this means real flame.
Regions, a normalized format, are one of many geometric elements that you can use in your decision making process. Other useful features of this kind are aspect ratios, moments, etc.

In short, you should do the following:
1 - Find places similar to fire.
2 - Trace the place of the candidate in all planes R, G, B.

3 - Extract a function (I suggest moments) in each plane.
4 - Generate a vector function
5 - Serve SVM with this vector

I hope you find this helpful.

Answer 2

Yes, so now your task is to create a .txt file with data on each image that you are going to process. Those that are true will be indicated by the +1 symbol after the set of functions and end with -1 and those that are false images of lights start with -1 and then with sets of functions and end with -1 before starting with the set of functions of the following image it will be a tedious job, but I'm sure ul controls this and finally save this file with the extension .train and not .txt so your workout file name will be filename.train