Fire scene residue trace identification method and system

Abstract

The invention relates to the technical field of general image data processing or generation, in particular to a fire scene residue trace identification method and system, and the method comprises the steps: obtaining previous images of a fire scene, carrying out the preprocessing of the obtained previous images of the fire scene, and obtaining an optimized image data set; importing the optimized image data set into a preset automatic identification model, executing target task selection, and obtaining a target task weight; on the basis of the target task weight, executing weighted summation of a loss value; executing back propagation, and finely adjusting the preset automatic identification model; and based on the preset automatic identification model after fine adjustment, executing identification. According to the method provided by the invention, under the condition that only a small amount of data exists, the target object can be successfully and accurately segmented by the model; and meanwhile, the fine adjustment of parameters by each back propagation is more accurate by using a loss function, a better final result is obtained, and the method has good universality and can convert an identification and segmentation target into traces of other fire scene residues.

Description

Technical Field

[0001] The present application relates to the general field of image data processing or generation technology, and in particular to a method for identifying traces of residues at a fire scene, and also to a system. Background Art

[0002] Traditionally, fire investigators have to observe and analyze fire scenes on site, often relying on experience and judgment. However, fire scenes are usually dirty and messy, and there are many types of items. If only manpower is relied upon, the overall identification of fire residues will be time-consuming and labor-intensive, which will restrict the efficiency of post-disaster residue identification and even fire cause determination.

[0003] With the development of computer science and the emergence of deep learning technology, deep learning models have become a common tool for automatic recognition. Common models include CNN models such as YOLO and Mask R-CNN, and segmentation models such as Swin Transformer.

[0004] However, deep learning-based models generally adopt a supervised training method. Since a large number of parameters need to be fine-tuned through back propagation, a large amount of labeled data is required, usually tens of thousands of sets of data. In real-life scenarios, there are limited images of fire scenes, and the number of images with metal plates in them is even less than the number required for fine-tuning. Therefore, if there is not enough data training or fine-tuning, the recognition and segmentation of traces of fire scene residues will be poor, the position of the metal plate cannot be accurately located, and it may even mislead the investigation. Summary of the invention

[0005] Through research, the inventor discovered that how to use limited image data to enable image recognition to accurately identify and locate traces of fire residues, especially metal sheets, has become an urgent problem that needs to be solved. If this goal can be achieved, it will save a lot of time and energy and improve the work efficiency of fire investigators.

[0006] The purpose of this application is to provide a method and system for identifying traces of residues at fire scenes. By combining a preset automatic recognition model with a semi-supervised training method, the technical problem that the prior art is unable to provide a method and system for accurately identifying and locating traces of residues at fire scenes using limited image data is solved.

[0007] According to one aspect of the present application, a method for identifying traces of residues at a fire scene is provided, and the method is executed by a processor, comprising:

[0008] Acquire past images of the fire scene, pre-process the acquired past images of the fire scene, and obtain an optimized image data set;

[0009] Import the optimized image dataset into the preset automatic recognition model, perform target task selection, and obtain the target task weight;

[0010] Based on the target task weight, perform weighted sum loss value;

[0011] Perform back propagation to fine-tune the preset automatic recognition model;

[0012] Perform recognition based on the fine-tuned preset automatic recognition model.

[0013] In some embodiments, the target task includes at least a training task and a reasoning task.

[0014] In some embodiments, when an inference task is selected, the optimized image data set is imported into a segmentation prediction module of a preset automatic recognition model, the position of the metal plate in the optimized image data set is segmented and identified, and the optimized image data set is saved. After saving, the execution is terminated.

[0015] In some embodiments, when a training task is selected, semi-supervised training of the reconstruction task and the segmentation task is performed; when the maximum number of training rounds is reached, the optimal weights in the training are saved and the execution is terminated.

[0016] In some embodiments, when the aforementioned training task has not reached the maximum training rounds, the optimized image data set is imported into the encoder for forward propagation, and the intermediate results obtained are alternately imported into the decoder of the unsupervised reconstruction task branch for continued forward propagation, and into the segmentation head of the supervised segmentation task branch for forward propagation.

[0017] In some embodiments, the intermediate result is passed through a loss function module of a preset automatic recognition model to compare the predicted image with the label image to calculate at least a cross entropy loss value, a patchwise loss value, and a Dice Loss loss value.

[0018] According to another aspect of the present application, a system for identifying traces of residues at a fire scene is provided, the system comprising a processor and further comprising:

[0019] An acquisition module, the acquisition module is used to acquire past images of the fire scene, pre-process the acquired past images of the fire scene, and obtain an optimized image data set;

[0020] A selection module, the selection module is used to import the optimized image data set into a preset automatic recognition model, perform target task selection, and obtain the target task weight;

[0021] A summation module, the summation module is used to perform a weighted summation loss value based on the target task weight;

[0022] An execution module, wherein the execution module is used to perform back propagation and fine-tune a preset automatic recognition model;

[0023] The recognition module is used to perform recognition based on a preset automatic recognition model after fine-tuning.

[0024] In some embodiments, the processor is data connected to the acquisition module, the selection module, the summation module, the execution module and the identification module.

[0025] Compared with the prior art, the present application has the following beneficial effects: the method of the present application can successfully enable the model to accurately segment the target object with only a small amount of data; at the same time, the loss function is used to make the fine-tuning of the parameters for each back propagation more accurate, thereby achieving a better final result. In addition, the method has good versatility and can convert the identification and segmentation targets into traces of other fire scene residues. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0027] Figure 1 is a flow chart of the identification method of the present application;

[0028] Figure 2 It is the structural diagram of the identification system of this application. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments.

[0030] This embodiment provides a method and system for identifying traces of fire scene residues. Figure 1-2 Describe them together.

[0031] Application Overview

[0032] The identification of fire scene residues is done by fire investigators through on-site observation and analysis, often based on experience and judgment. However, fire scenes are usually dirty and messy, and there are many kinds of items. If only relying on manpower, the overall identification of residues will be time-consuming and labor-intensive, which restricts the efficiency of post-disaster residue identification and even fire cause determination. Therefore, how to accurately identify and locate fire scene residues, especially metal pieces, has become an urgent problem to be solved. If this goal can be achieved, the work efficiency of fire investigators will be improved.

[0033] The present application aims to provide a method and system for identifying traces of fire scene residues. By combining a preset automatic recognition model with a semi-supervised training method, the problem that the existing technology is unable to provide limited image data for accurate identification and positioning of traces of fire scene residues is solved.

[0034] Exemplary Methods

[0035] The present embodiment provides a method for identifying traces of residues at a fire scene. The method is executed by a processor and includes: acquiring past images of the fire scene, preprocessing the acquired past images of the fire scene, and obtaining an optimized image data set; importing the optimized image data set into a preset automatic recognition model, performing target task selection, and obtaining target task weights; performing weighted sum loss value based on the target task weights; performing back propagation and fine-tuning the preset automatic recognition model; and performing recognition based on the fine-tuned preset automatic recognition model.

[0036] This method performs weight and loss value calculations on the optimized image dataset. By back-propagating to the preset automatic recognition model, it can accurately locate the metal pieces at the fire scene. It also has good versatility and can convert the recognition and segmentation targets into traces of other fire scene residues.

[0037] Below, each step will be described in detail.

[0038] First, obtain the previous images of the fire scene, preprocess the obtained previous images of the fire scene, and obtain the optimized image data set. In detail, the original data of the previous images of the fire scene are passed into and run the image training preprocessing module of the preset automatic recognition model, and the images are uniformly converted into a general format that can be input, that is, npy weight files; further, the image size is changed to 512×512 pixels; then, the color fire scene image is converted into a grayscale image, the average value of each pixel is calculated and the value of each pixel is normalized; further, part of the image is selectively flipped and rotated to improve the generalization ability of the preset automatic recognition model, and a set of preprocessed and enhanced image data is obtained, that is, an optimized image data set, which can be directly input into any model for training or reasoning, and in this implementation, it is a preset automatic recognition model.

[0039] Next, the optimized image data set is imported into a preset automatic recognition model, target task selection is performed, and target task weights are obtained. Specifically, in some implementations, the target task includes at least a training task and an inference task.

[0040] When the reasoning task is selected, the optimized image data set is imported into the segmentation prediction module of the preset automatic recognition model, the metal plate position in the optimized image data set is segmented and identified, and the optimized image data set is saved. After saving, the execution ends.

[0041] When a training task is selected, semi-supervised training of the reconstruction task and the segmentation task is performed; when the maximum number of training rounds is reached, the optimal weights in the training are saved and the execution ends.

[0042] It is understandable that when the aforementioned training task has not reached the maximum training rounds, the optimized image data set is imported into the encoder for forward propagation, and the intermediate results are alternately imported into the MAE decoder of the unsupervised reconstruction task branch for continued forward propagation, and the MLA segmentation head of the supervised segmentation task branch for forward propagation. Then, the intermediate results are passed through the loss function module of the preset automatic recognition model to compare the predicted image with the label image to at least calculate the cross entropy loss value, patchwise loss value and Dice Loss loss value. It is understandable that in some possible implementations, the training rounds are judged, and if the training rounds of a specific interval are reached, the model weights are saved; if the maximum training rounds are not reached, the semi-supervised training of the reconstruction task and the segmentation task is cyclically executed until the maximum training rounds are reached, and the execution is terminated.

[0043] At the same time, it should be noted that the cross entropy function is particularly suitable for processing multi-class classification tasks, which is conducive to the training model to correctly identify metal plates from various objects; the patch calculation loss function can calculate the batch with the largest loss, and then increase its weight through the temperature coefficient, while ensuring that the overall loss does not increase, which not only ensures faster training speed, but also improves training results; the Dice calculation loss function can be used to measure the overlapping area between the predicted image and the label image, forcing the model to segment the metal plate area as accurately as possible.

[0044] At the same time, it should be noted that semi-supervised training is divided into unsupervised training without labels and supervised training with labeled data. The data used in unsupervised training is the original image of the fire scene, but there is no corresponding label. The training method is to randomly cover a number of image blocks on the original image, and the covered area accounts for 75%, forcing the preset automatic recognition model to obtain a deep understanding of the image, so as to predict an image as close to the original image as possible, and then input the model forward propagation, and reconstruct the covered area of ​​the original image based on the remaining image. Finally, the loss function is calculated for the reconstructed area and the area corresponding to the original image, and the parameters are fine-tuned by back propagation.

[0045] Secondly, based on the aforementioned target task weights, perform weighted summation of the aforementioned loss values; perform back propagation and fine-tune the preset automatic recognition model.

[0046] Finally, recognition is performed based on the fine-tuned preset automatic recognition model.

[0047] In order to better understand the present application, the present embodiment further provides a system for identifying traces of residues at a fire scene, the system comprising a processor and further comprising:

[0048] An acquisition module, the acquisition module is used to acquire past images of the fire scene, pre-process the acquired past images of the fire scene, and obtain an optimized image data set;

[0049] A selection module, which is used to import the optimized image data set into a preset automatic recognition model, perform target task selection, and obtain target task weights;

[0050] The summation module is used to perform weighted summation loss value based on the target task weight;

[0051] An execution module, which is used to perform back propagation and fine-tune the preset automatic recognition model;

[0052] Recognition module,The recognition module is used to perform recognition based on a fine-tuned preset automatic recognition model.

[0053] Among them, the processor data connection acquisition module, selection module, summation module, execution module and identification module.

[0054] Exemplarily, the system can be divided into one or more modules, one or more modules are stored in the memory and executed by the processor to complete the present application. One or more modules can be a series of computer program instruction segments that can complete specific functions, and the instruction segments are used to describe the execution process of the computer program in the terminal device. For example, the system can be divided into an acquisition module, a selection module, a summation module, an execution module, and an identification module, and the specific functions of each module are as described above.

[0055] The aforementioned terminal device may be a computing device such as a desktop computer, a notebook, a PDA, and a cloud server. The terminal device may include, but is not limited to, a processor and a memory. Those skilled in the art will appreciate that the examples of the terminal device do not constitute a limitation on the terminal device.

[0056] Meanwhile, the aforementioned processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The aforementioned processor is the control center of the identification system, and uses various interfaces and lines to connect various parts of the entire terminal device.

[0057] The above shows and describes the basic principles, main features and advantages of the present application. For those skilled in the art, it is obvious that the present application is not limited to the details of the above-mentioned exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or basic features of the present application.

[0058] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for identifying traces of residues at a fire scene, the method being executed by a processor, characterized in that: include: Acquire past images of the fire scene, pre-process the acquired past images of the fire scene, and obtain an optimized image data set; Import the optimized image dataset into the preset automatic recognition model, perform target task selection, and obtain the target task weight; Based on the target task weight, perform weighted sum loss value; Perform back propagation to fine-tune the preset automatic recognition model; Perform recognition based on the fine-tuned preset automatic recognition model.

2. The identification method according to claim 1, characterized in that: The target task includes at least a training task and a reasoning task.

3. The identification method according to claim 2, characterized in that: When the reasoning task is selected, the optimized image data set is imported into the segmentation prediction module of the preset automatic recognition model, the metal plate position in the optimized image data set is segmented and identified, and the optimized image data set is saved. After saving, the execution ends.

4. The identification method according to claim 2, characterized in that: When a training task is selected, semi-supervised training of the reconstruction task and the segmentation task is performed; when the maximum number of training rounds is reached, the optimal weights in the training are saved and the execution ends.

5. The identification method according to claim 4, characterized in that: When the aforementioned training task has not reached the maximum number of training rounds, the optimized image dataset is imported into the encoder for forward propagation, and the intermediate results are alternately imported into the decoder of the unsupervised reconstruction task branch for continued forward propagation and the segmentation head of the supervised segmentation task branch for forward propagation.

6. The identification method according to claim 5, characterized in that: The intermediate results are passed through the loss function module of the preset automatic recognition model to compare the predicted image with the label image to at least calculate the cross entropy loss value, patchwise loss value and Dice Loss loss value respectively.

7. A fire scene residue trace identification system, the system comprising a processor, characterized in that: Also includes: An acquisition module, the acquisition module is used to acquire past images of the fire scene, pre-process the acquired past images of the fire scene, and obtain an optimized image data set; A selection module, the selection module is used to import the optimized image data set into a preset automatic recognition model, perform target task selection, and obtain the target task weight; A summation module, the summation module is used to perform a weighted summation loss value based on the target task weight; An execution module, wherein the execution module is used to perform back propagation and fine-tune a preset automatic recognition model; The recognition module is used to perform recognition based on a preset automatic recognition model after fine-tuning.

8. The system according to claim 7, characterized in that The processor is data connected to the acquisition module, the selection module, the summing module, the execution module and the identification module.

Images