Camouflage object detection method based on deep learning uncertain form

Abstract

The invention provides a camouflage object detection method based on a deep learning uncertain form. The method comprises the following steps: acquiring the spectral wavelength and spectral energy of flames when hydrogen and chlorine are combusted in different proportions through a spectrograph; performing data preprocessing on the spectral wavelength and the spectral energy; inputting the preprocessed spectral wavelength and spectral energy into a constructed deep learning network model for training; in the training process, the training round and the learning rate are adjusted until an optimal weight file is obtained; and performing real-time prediction on spectral wavelength and spectral energy output by the spectrometer interface through the optimal weight file and the deep learning network, and regulating and controlling the ratio of chlorine to hydrogen according to a prediction result. The method provided by the invention effectively overcomes the defects of a manual observation method and a distributed control system, and further improves the preparation efficiency of hydrogen chloride.

Description

Technical Field

[0001] The present invention relates to the field of industrial preparation technology, and in particular to a camouflaged object detection method based on deep learning uncertain morphology. Background Art

[0002] The industrial preparation of hydrochloric acid is a complex process. Many chemical plants mainly obtain it through the complete combustion of chlorine and hydrogen in a synthesis furnace. During the chemical reaction process of the synthesis furnace, the flame exists in the form of plasma and the flame color is pale. This uncertain state and weak visual perception phenomenon will greatly affect the final quality of the product when the ratio of chlorine and hydrogen is subsequently adjusted. Therefore, it is very important to balance the ratio of chlorine and hydrogen.

[0003] There are two main control methods for balancing the two gases: (1) manual observation, which determines the ratio of chlorine and hydrogen by observing the color of the flame burning in the synthesis furnace. Since the form of the substance in the synthesis furnace combustion process is difficult to visually perceive, this method is inefficient and the product quality varies. (2) Distributed control system, which can automatically adjust the pressure, flow and other parameter setting values ​​to the ratio of chlorine and hydrogen, which can effectively improve the quality of the product. However, the equipment is expensive and the maintenance process is cumbersome. The visual perception equipment needs to be cleaned regularly to ensure the normal operation of the equipment. Summary of the invention

[0004] In view of this, the present invention provides a camouflaged object detection method based on deep learning uncertain morphology to solve the above problems.

[0005] The present invention provides a camouflaged object detection method based on deep learning uncertain morphology, comprising: obtaining the spectral wavelength and spectral energy of the flame when hydrogen and chlorine are burned in different ratios through a spectrometer; performing data preprocessing on the spectral wavelength and spectral energy; inputting the preprocessed spectral wavelength and spectral energy into a constructed deep learning network model for training; during the training process, adjusting the training rounds and the learning rate until an optimal weight file is obtained; performing real-time prediction of the spectral wavelength and spectral energy output by the spectrometer interface through the optimal weight file and the deep learning network, and regulating the ratio of chlorine and hydrogen according to the prediction result.

[0006] In another implementation of the present invention, the data preprocessing includes: converting the spectral wavelength and the spectral energy into a deep learning type.

[0007] In another implementation of the present invention, it also includes: constructing the deep learning network model through one-dimensional convolution neural operation, maximum pooling operation, full connection operation, activation function, Dropout operation and Flatten operation.

[0008] In another implementation of the present invention, it also includes: verifying the performance of the deep learning type model through evaluation indicators, and the evaluation indicators include corrected determination coefficient, mean square error loss function and root mean square error.

[0009] In another implementation of the present invention, the mean square error is expressed as:

[0010]

[0011] Among them, f(x) represents the predicted value, y represents the true value label; the closer the mean square error is to 1, the more accurate the prediction result is.

[0012] In another implementation of the present invention, the corrected determination coefficient is expressed as:

[0013]

[0014] Among them, y i represents the true observed value, Represents the prediction result, It represents the average value of the real observations; the closer the corrected determination coefficient is to 1, the more accurate the prediction result is.

[0015] In another implementation of the present invention, the root mean square error is expressed as:

[0016]

[0017] Among them, y i represents the true observed value, Represents the prediction result; the closer the value of the root mean square error is to 0, the smaller the error between the prediction result and the true value label is.

[0018] The camouflage object detection method based on deep learning of uncertain morphology of the present invention has a control process visualization effect, and can observe the changes in flame color in the synthesis furnace in real time; it uses a neural network to detect the combustion state of chlorine and hydrogen with high efficiency and accuracy; it analyzes the ratio and combustion state of hydrogen and chlorine through spectral wavelength and spectral energy, which can solve the defects in visual perception; in complex scenes, it can replace humans to make better and more accurate judgments and analysis problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. By reading the detailed description of the following implementations, the advantages and benefits of the solutions become clear to those skilled in the art. The drawings are only used to illustrate the preferred implementations and are not to be considered as limiting the present invention.

[0020] In the attached picture:

[0021] Figure 1 The figure is a flowchart of a method for detecting camouflaged objects based on uncertain morphology based on deep learning according to an embodiment of the present invention.

[0022] Figure 2 FIG. 4 is a schematic diagram of spectral wavelength and spectral energy according to an embodiment of the present invention.

[0023] Figure 3 A schematic diagram of a training data set according to an embodiment of the present invention.

[0024] Figure 4 A schematic diagram of a deep learning network model according to an embodiment of the present invention.

[0025] Figure 5 This is a convergence diagram of 30 training rounds according to an embodiment of the present invention.

[0026] Figure 6 This is a convergence diagram of 100 training rounds according to an embodiment of the present invention.

[0027] Figure 7 This is a convergence diagram of 300 training rounds according to an embodiment of the present invention.

[0028] Figure 8 This is a convergence diagram of 1000 training rounds according to an embodiment of the present invention.

[0029] Fig. 9 The figure is a schematic diagram of prediction results according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] For ease of understanding, before describing in detail the specific embodiments of the present invention, an existing technology of a camouflaged object detection method based on deep learning uncertain morphology of the present invention is first exemplarily described.

[0031] Since the phenomena produced by the chemical combustion process of chlorine and hydrogen have uncertainty in perception in the human visual system, the plasma state and color of the flame are a dynamically changing process. This uncertainty phenomenon is similar to the camouflage phenomenon. Therefore, this phenomenon is called industrial camouflage phenomenon. This industrial camouflage phenomenon largely determines the quality of the product. Capturing this uncertainty phenomenon and analyzing and classifying it can well solve the deficiencies in the existing production process.

[0032] In addition, the development of deep learning and camouflage target detection technology has enabled many uncertain objects or phenomena in ecology, agriculture, and industry to be displayed through significant methods. Therefore, the use of deep learning technology to detect flame combustion and determine the ratio of gas concentration in the synthesis furnace can not only effectively improve the efficiency of industrial preparation of hydrochloric acid, but also effectively make up for the shortcomings of manual observation and control, and at the same time make the production process more intelligent.

[0033] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be described clearly and in detail below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in the field based on the embodiments in the embodiments of the present invention should fall within the scope of protection of the embodiments of the present invention.

[0034] Figure 1 A flowchart of a method for detecting camouflaged objects based on deep learning uncertain morphology is provided in an embodiment of the present invention, such as Figure 1 As shown, this embodiment mainly includes:

[0035] S101. Obtain the spectral wavelength and spectral energy of the flame when hydrogen and chlorine are burned in different ratios through a spectrometer.

[0036] For example, Figure 2 As shown, the spectral wavelength and spectral energy under three different states collected by the spectrometer. The spectral wavelength is represented by the horizontal axis, and the spectral energy is represented by the vertical axis. The spectral wavelength and spectral energy distribution corresponding to the color of the synthesis furnace flame change more obviously in the position along the way.

[0037] The collected spectral wavelengths and spectral energies include:

[0038] The spectral wavelength and spectral energy of the flame when the ratio of hydrogen and chlorine in the synthesis furnace is normal, the spectral wavelength and spectral energy of the flame when the chlorine in the synthesis furnace becomes less, and the spectral wavelength and spectral energy of the flame when the hydrogen in the synthesis furnace becomes less.

[0039] S102: performing data preprocessing on the spectral wavelength and spectral energy.

[0040] S103, inputting the preprocessed spectral wavelength and spectral energy into the constructed deep learning network model for training.

[0041] S104. During the training process, adjust the training rounds and the learning rate until the optimal weight file is obtained.

[0042] For example, Figure 5 ,6 As shown in , 7 and 8, the training rounds are 30 rounds, 100 rounds, 300 rounds and 1000 rounds respectively. Different training rounds and learning rates are set to obtain the best weight file. After multiple rounds of training, the best weight file is saved. The best learning rate obtained in the present invention is 0.003, and the best number of training rounds is determined according to the number of samples in the data set.

[0043] S105. Predict the spectral wavelength and spectral energy output by the spectrometer interface in real time through the optimal weight file and the deep learning network, and adjust the ratio of chlorine and hydrogen according to the prediction results.

[0044] For example, by calling the data transmission interface of the spectrometer to collect spectral wavelength and spectral energy, the collected data is saved in the form of a table file to the root directory of the prediction data set pointed to by the prediction file, so as to achieve real-time prediction. Before prediction, the real-time collected spectral energy needs to be preprocessed and the collected data is converted into a deep learning type (torch.tensor(spectral energy, torch.float32).to('cuda')). By calling the network model and loading the optimal weight file, the result of the real-time collected data can be predicted. By predicting the real-time collected spectral energy, it can be concluded what type of spectrum the spectral energy belongs to.

[0045] The camouflage object detection method based on deep learning of uncertain morphology of the present invention has a control process visualization effect, and can observe the changes in flame color in the synthesis furnace in real time; it uses a neural network to detect the combustion state of chlorine and hydrogen with high efficiency and accuracy; it analyzes the ratio and combustion state of hydrogen and chlorine through spectral wavelength and spectral energy, which can solve the defects in visual perception; in complex scenes, it can replace humans to make better and more accurate judgments and analysis problems.

[0046] In another implementation of the present invention, the data preprocessing includes: converting the spectral wavelength and the spectral energy into a deep learning type.

[0047] Exemplarily, the processing method for producing a training data set includes: using a connected spectrometer device close to the window of the combustion area of ​​the synthesis furnace; obtaining the real-time flame color through the sensing probe of the spectrometer; using the supporting software of the spectrometer to parse out the spectral wavelength and spectral energy; the spectral wavelength and spectral energy parsed by the supporting software are stored in the form of a table file.

[0048] Since the type of sample cannot be known when predicting unknown data, there is no need to set the type of sample in the table file when the collected data is used as a prediction data set. The first row of data in the table represents spectral wavelength; the first column of data in the table represents the sample type, which is used to distinguish the normal state of flame combustion in the synthesis furnace from the state after the concentration of hydrogen and chlorine is changed; the data from the second column to the last column and the second row to the last row in the table represent spectral energy.

[0049] like Figure 3 As shown in the figure, λ1…λn represents the spectral wavelength, 1 represents the sample type, and XXX represents the spectral energy. The sample categories of the data in the training data set are represented by digital numbers. The data category sampled when the hydrogen and chlorine in the synthesis furnace are burned in a normal ratio is represented as 1; the data category sampled when the chlorine in the synthesis furnace becomes less is represented as 2; the data category sampled when the hydrogen in the synthesis furnace becomes less is represented as 3; 200 data of each category are sampled and stored in the order of category numbers. The more data samples are collected, the better the performance of the network model after training.

[0050] Before implementing network model training, the data in the training data set needs to be preprocessed. The preprocessing methods include:

[0051] (1) Convert the sample type (true value label) in the training dataset into a deep learning type (torch.tensor('sample type', torch.float32).to('cuda')).

[0052] (2) The spectral energy of the training data set is also converted into deep learning type.

[0053] In another implementation of the present invention, it also includes: constructing the deep learning network model through one-dimensional convolution neural operation, maximum pooling operation, full connection operation, activation function, Dropout operation and Flatten operation.

[0054] For example, Figure 4 As shown in the figure, the deep learning network model includes: feature extraction, pooling operation processing, Flatten operation full connection processing (including full connection operation, activation function and Dropout operation). The convolutional neural network is used to extract feature information from the preprocessed data. The network model has 15 stages, represented by Stage 1 to Stage 15.

[0055] Stage 1 is represented as a convolutional layer, which is used to preprocess the feature extraction of input information. The input channel of Stage 1 is 1, the output channel is 16, the convolution kernel size is set to 13×13, and the stride is set to 2.

[0056] Stage 2 is represented as a convolutional layer, which is used to extract features from the feature map Stage 1. The input channel of Stage 2 is 16, the output channel is 32, the convolution kernel size is set to 11×11, and the stride is set to 2.

[0057] Stage3 is represented as a pooling layer, which is used to reduce the dimension of the feature map Stage2 and compress the feature map, simplify the network complexity, etc. The convolution kernel size of Stage3 is set to 2×2, and the step size is set to 1.

[0058] Stage 4 is represented as a convolutional layer, which is used to extract features from the feature map Stage 3. The input channel of Stage 4 is 32, the output channel is 64, the convolution kernel size is set to 9×9, and the stride is set to 1.

[0059] Stage 5 is represented as a convolutional layer, which is used to extract spectral features from feature map Stage 4. The input channel of Stage 5 is 64, the output channel is 64, the convolution kernel size is set to 7×7, and the stride is set to 2.

[0060] Stage 6 is represented as a pooling layer, which is used to reduce the dimension of the feature map Stage 5 and compress the feature map, simplify the network complexity, etc. The convolution kernel size of Stage 7 is set to 2×2 and the step size is set to 1.

[0061] Stage 7 is represented as a convolutional layer, which is used to extract features from the feature map Stage 6. The input channel of Stage 8 is 64, the output channel is 128, the convolution kernel size is set to 5×5, and the stride is set to 1.

[0062] Stage 8 is represented as a convolutional layer, which is used to extract spectral features from feature map Stage 7. The input channel of Stage 8 is 128, the output channel is 256, the convolution kernel size is set to 3×3, and the stride is set to 2.

[0063] Stage9 is represented as a pooling layer, which is used to reduce the dimension of the feature map Stage8 and compress the feature map, simplify the network complexity, etc. The convolution kernel size of Stage7 is set to 2×2 and the stride is set to 2.

[0064] Stage 10 is represented by the Flatten operation, which is used to flatten the input tensor and merge all dimensions except the batch dimension and the channel dimension into one dimension.

[0065] Stage11 is represented as a fully connected layer, which is used to convert the feature information processed by the previous stage into the final classification. The input feature of the fully connected operation is 1280 and the output feature is 512. The ReLU activation function is used for smoothing, and the parameter of the Dropout operation is set to 0.2.

[0066] Stage 12 is represented as a fully connected layer. The input feature of the fully connected operation is 512, and the output feature is 256. The ReLU activation function is used for smoothing, and the parameter of the Dropout operation is set to 0.2.

[0067] Stage 13 is represented as a fully connected layer. The input feature of the fully connected operation is 256 and the output feature is 64. The ReLU activation function is used for smoothing, and the parameter of the Dropout operation is set to 0.2.

[0068] Stage 14 is represented as a fully connected layer. The input feature of the fully connected operation is 64, and the output feature is 32. The ReLU activation function is used for smoothing, and the parameter of the Dropout operation is set to 0.2.

[0069] Stage 15 is represented as a separate fully connected operation with an input feature of 32 and an output feature of 1.

[0070] Finally, Stage 1 to Stage 15 are encapsulated as a whole through Sequential operations to complete the design of the entire network model.

[0071] In another implementation of the present invention, it also includes: verifying the performance of the deep learning type model through evaluation indicators, and the evaluation indicators include corrected determination coefficient, mean square error loss function and root mean square error.

[0072] In another implementation of the present invention, the mean square error is expressed as:

[0073]

[0074] Among them, f(x) represents the predicted value, y represents the true value label; the closer the mean square error is to 1, the more accurate the prediction result is.

[0075] For example, the mean squared error loss function (MSE) is used to calculate the loss value during the training process. The MSE function has the characteristics of smooth and continuous curves and the curve can be differentiated everywhere, which is convenient for using the gradient descent algorithm. Moreover, as the error decreases, the gradient also decreases, which is conducive to convergence. Even if a fixed learning rate is used, it can converge to the minimum value quickly.

[0076] In another implementation of the present invention, the corrected determination coefficient is expressed as:

[0077]

[0078] Among them, y i represents the true observed value, Represents the prediction result, It represents the average value of the real observations; the closer the corrected determination coefficient is to 1, the more accurate the prediction result is.

[0079] For example, the adjusted R-squared, also known as R 2 The score is calculated by determining the degree of fit of the linear regression line to calculate the difference between the independent variable (the independent variable is equivalent to the true value label) and the dependent variable (the dependent variable is equivalent to the predicted result).

[0080] In another implementation of the present invention, the root mean square error is expressed as:

[0081]

[0082] Among them, y i represents the true observed value, Represents the prediction result; the closer the value of the root mean square error is to 0, the smaller the error between the prediction result and the true value label is.

[0083] For example, the root mean square error (RMSE) refers to the square root of the ratio of the square of the deviation between the predicted result and the true observed value to the number of observations.

[0084] In another implementation of the present invention, a stochastic gradient descent (SGD) method is used to optimize the learning rate of the network model so as to achieve better convergence efficiency during network training.

[0085] In another implementation of the present invention, Fig. 9 As shown, in order to better display the prediction results, sample data in three different states are collected to display the prediction results. The discrete points at the horizontal axis label 1 represent the prediction results when the hydrogen and chlorine are in a normal ratio; the discrete points at the horizontal axis label 2 represent the prediction results when the chlorine becomes less; the discrete points at the horizontal axis label 3 represent the prediction results when the hydrogen becomes less.

[0086] Furthermore, the closer the prediction result is to 1, the more normal the ratio of hydrogen and chlorine in the synthesis furnace is; when the prediction result is close to 2, it indicates that the chlorine in the synthesis furnace is decreasing; if the prediction result is close to 3, it indicates that the hydrogen in the synthesis furnace is decreasing.

[0087] The experimental results show that the accuracy of the predicted results of the three states is 0.985.

[0088] In another aspect of the present invention, an electronic device includes: a processor, a memory, a communication bus, and a communication interface.

[0089] in:

[0090] The processor, memory and communication interface communicate with each other through a communication bus.

[0091] Communication interface, used to communicate with other electronic devices or servers.

[0092] The processor is used to execute the program, and specifically can execute the steps of any one of the camouflaged object detection methods based on deep learning and uncertain morphology in the above embodiments.

[0093] Specifically, the program may include program codes including computer operation instructions.

[0094] The processor may be a central processing unit (CPU) or a graphics processing unit (GPU). The one or more processors included in the smart device may be processors of the same type, such as one or more CPUs, or processors of different types, such as one or more CPUs and one or more GPUs.

[0095] The memory is used to store programs. The memory may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory.

[0096] The program can be specifically used to enable the processor to execute to implement the steps of any of the camouflaged object detection methods based on deep learning uncertain forms described in the embodiments. The specific implementation of each step in the program can refer to the corresponding descriptions in the steps and units executed by any of the camouflaged object detection methods based on deep learning uncertain forms in the above steps, which will not be repeated here. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the above-described devices and modules can refer to the corresponding process description in the aforementioned method embodiments.

[0097] The method according to the embodiment of the present invention may be implemented in a server equipped with a central processing unit (CPU) and an image processing unit (GPU).

[0098] So far, specific embodiments of the present invention have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims can be performed in a different order and still achieve the desired results. In addition, the sample types corresponding to the training data sets in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results.

[0099] It should be noted that all directional indications in the embodiments of the present invention (such as up, down, left, right, back, etc.) are only used to explain the relative position relationship between the components in a certain specific order (as shown in the accompanying drawings). If the specific order changes, the directional indication will also change accordingly.

[0100] In the description of the present invention, the terms "first" and "second" are only used to facilitate the description of different components or names, and cannot be understood as indicating or implying a sequential relationship, relative importance, or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features.

[0101] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0102] It should be noted that although the specific embodiments of the present invention are described in detail in conjunction with the accompanying drawings, it should not be understood as limiting the scope of protection of the present invention. Within the scope described in the claims, various modifications and variations that can be made by those skilled in the art without creative work still belong to the scope of protection of the present invention.

[0103] The examples of the embodiments of the present invention are intended to concisely illustrate the technical features of the embodiments of the present invention so that those skilled in the art can intuitively understand the technical features of the embodiments of the present invention, and are not intended to be improper limitations of the embodiments of the present invention.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A camouflaged object detection method based on deep learning uncertain morphology, characterized in that: include: The spectrum wavelength and spectrum energy of the flame when hydrogen and chlorine are burned at different ratios are obtained by using a spectrometer; Performing data preprocessing on the spectral wavelength and spectral energy; The preprocessed spectral wavelength and spectral energy are input into the constructed deep learning network model for training; During the training process, adjust the training rounds and learning rate until the best weight file is obtained; The optimal weight file and the deep learning network are used to predict the spectral wavelength and spectral energy output by the spectrometer interface in real time, and the ratio of chlorine and hydrogen is regulated according to the prediction results.

2. The method according to claim 1, characterized in that The data preprocessing includes: The spectral wavelength and the spectral energy are converted into deep learning types.

3. The method according to claim 1, characterized in that Also includes: The deep learning network model is constructed through one-dimensional convolution neural operation, maximum pooling operation, full connection operation, activation function, Dropout operation and Flatten operation.

4. The method according to claim 1, characterized in that: Also includes: The performance of the deep learning type model is verified by evaluation indicators, including the corrected determination coefficient, the mean square error loss function and the root mean square error.

5. The method according to claim 4, characterized in that The mean square error is expressed as: Among them, f(x) represents the predicted value and y represents the true value label; The closer the mean square error is to 1, the more accurate the prediction result is.

6. The method according to claim 4, characterized in that The corrected coefficient of determination is expressed as: Among them, y i represents the true observed value, Represents the prediction result, represents the mean of the true observations; The closer the corrected determination coefficient is to 1, the more accurate the prediction result is.

7. The method according to claim 4, characterized in that The root mean square error is expressed as: Among them, y i represents the true observed value, Indicates the prediction result; The closer the value of the root mean square error is to 0, the smaller the error between the prediction result and the true value label is.

Images