Image classification method and device based on super-dimensional calculation, medium and product

Abstract

The invention provides an image classification method and device based on super-dimensional calculation, a medium and a product. The method comprises the following steps: extracting target texture features from an infrared image by using a local binary pattern; performing feature dimension reduction on the target texture features through a principal component analysis method to obtain one-dimensional structure feature information of the target; performing super-dimensional coding on the one-dimensional structure feature information of the target by using an efficient multivariate coding method to form a super-dimensional vector; constructing a super-dimensional vector data set by using the obtained super-dimensional vectors; constructing a target classification and recognition model based on a binary full-connection neural network; training a target classification recognition model by using the super-dimensional vector data set; and performing target recognition by using the trained target classification recognition model to obtain a target category. Target recognition is completed through combination of feature extraction, efficient super-dimensional coding and the binary full-connection neural network, the image classification recognition accuracy can be improved, and the method has the advantages of being low in super-dimensional coding complexity, high in model generalization ability and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target recognition, in particular to an image classification method and device based on hyper-dimensional computation, medium and product. BACKGROUND

[0002] At present, in the field of image classification and recognition, the direct processing objects based on hyper-dimensional computation are mainly small-sized images, and there is no better processing method for directly hyper-dimensional coding of large-sized images. Meanwhile, in image classification and recognition, hyper-dimensional coding is mainly carried out by using recording-type coding technology and multi-element-type coding technology, but these two coding methods are slightly complex, slow and low in efficiency. For the hyper-dimensional vector after hyper-dimensional coding, the mainstream method is to use Hamming distance or cosine distance calculation method to determine the category, and the classification and recognition rate is not high. SUMMARY

[0003] The present application aims to provide an image classification method and device based on hyper-dimensional computation, medium and product to solve the above problems.

[0004] In a first aspect, the present application provides an image classification method based on hyper-dimensional computation, comprising: extracting features of the infrared image by using a local binary pattern to obtain texture features of the target; performing feature dimension reduction on the texture features of the target by using a principal component analysis method to obtain one-dimensional structural feature information of the target; performing hyper-dimensional coding on the one-dimensional structural feature information of the target by using a high-efficiency multi-element-type coding method to form a hyper-dimensional vector; constructing a hyper-dimensional vector dataset from the obtained hyper-dimensional vector; constructing a target classification and recognition model based on a binary full-connection neural network; training the target classification and recognition model based on the binary full-connection neural network by using the hyper-dimensional vector dataset; sending the hyper-dimensional vector formed after feature extraction, feature dimension reduction and hyper-dimensional coding of the collected infrared image into the trained target classification and recognition model to perform target recognition, and obtaining the target category.

[0005] In a preferred embodiment, the hyper-dimensional coding on the one-dimensional structural feature information of the target by using the high-efficiency multi-element-type coding method to form a hyper-dimensional vector comprises: quantizing the feature values of the one-dimensional structural feature information of the target, and forming a hyper-dimensional vector by using structured density coding on the quantized feature values, wherein the dimension of the feature vector is G obtaining the hyper-dimensional vector corresponding to the g th feature value in the feature vector by using structured density coding , wherein DThe super-dimension coding dimension; The super-dimension vector located at the g The super-dimension vector located at the The permutation operation G - g The actual operation is a cyclic right shift G - g The bit; The numerical super-vector after the respective shift operation is bound by multiplication to obtain the final D-dimensional super-vector, denoted as: ; wherein, R represents the super-dimension vector based on the efficient multi-element coding method, represents a cyclic right shift G -1 bit, and the same applies; represents the multiplication operation.

[0006] In a preferred embodiment, the characteristic value quantization formula is as follows:

[0007] wherein, D represents the dimension of the super-dimension vector, U F represents the characteristic value of the one-dimensional structural feature information before quantization, U Q represents the quantized characteristic value, represents the rounding operation.

[0008] In a preferred embodiment, the structured density coding adopts bipolar coding or binary coding.

[0009] In a preferred embodiment, the structure of the target classification and recognition model based on the binary full connection neural network includes an input layer, a first full connection layer, a first binary activation function layer, a second full connection layer, a second binary activation function layer, and an output layer connected in sequence.

[0010] In a preferred embodiment, the calculation process of the full connection layer in the target classification and recognition model based on the binary full connection neural network includes: Deterministic binarization: the non-binary weight and input characteristic value are converted into 0 and 1 form through the Sign function; XOR operation is performed to obtain the XOR operation result; The number of features with the XOR operation result being 1, Popcount_xor, is counted, and the final result is the total number of features minus twice the Popcount_xor.

[0011] In a preferred embodiment, when training the target classification identification model based on the binary full connection neural network, the linear layer is used in combination with the Sign function to complete the forward propagation process, and the Clip function is used to replace the Sign function in the backward propagation process to ensure that the gradient can be calculated; in the network inference process, the Sign function is used to realize the exclusive or operation.

[0012] In a second aspect, the present application provides an electronic device, comprising: at least one processor; and a memory connected to the at least one processor in communication; The memory stores instructions executable by the at least one processor, and the at least one processor executes the instructions stored in the memory, so that the at least one processor executes the method.

[0013] In a third aspect, the present application provides a computer readable storage medium for storing instructions, when the instructions are executed, the method is realized.

[0014] In a fourth aspect, the present application provides a computer program product, when the computer program product is called by a computer, the computer executes the above-mentioned method.

[0015] In summary, the present application combines feature extraction, efficient hyper-encoding and binary full connection neural network to complete target identification, which can improve the image classification and recognition accuracy, and has the advantages of low hyper-encoding complexity, strong model generalization ability and the like. Specifically as follows: 1. The present application extracts effective features of the target by local binary pattern and principal component analysis method, which can solve the problem of large size image input being too large to be hyper-encoded.

[0016] 2. The present application uses a high-efficiency multivariate encoding method to improve the hyper-encoding speed, reduce the encoding calculation complexity, and at the same time, the precision loss is small.

[0017] 3. The binary full connection neural network of the present application uses exclusive or binary full connection operation, which can reduce the consumption of hardware resources and improve the inference speed, can meet the use of edge device based on FPGA under the condition of limited resources, and at the same time can improve the accuracy of the whole target identification.

[0018] 4. Compared with the traditional multivariate hyper-encoding method, the encoding speed of the method is faster, and the target classification accuracy can be improved by more than 3%. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A flow chart of an image classification method based on hyper-computing according to an embodiment of the present application.

[0020] Figure 2 A general block diagram of the efficient multi-type encoding method in the embodiment of the present application.

[0021] Figure 3 A flowchart of the eigenvalue quantization in the embodiment of the present application.

[0022] Figure 4 A schematic diagram of the structured density encoding in the embodiment of the present application.

[0023] Figure 5 A flowchart of the training and inference of the binary fully connected neural network in the embodiment of the present application.

[0024] Figure 6 A schematic diagram of the calculation process of the fully connected layer in the binary fully connected neural network in the embodiment of the present application.

[0025] Figure 7 A schematic diagram of the Sign binary activation function and its derivative in the embodiment of the present application.

[0026] Figure 8 A schematic diagram of the Clip binary activation function and its derivative in the embodiment of the present application.

[0027] Figure 9 A schematic diagram of the framework of the binary fully connected neural network in the embodiment of the present application.

[0028] Figure 10 A schematic diagram of the structure of an electronic device proposed in the embodiment of the present application. DETAILED DESCRIPTION

[0029] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0030] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative labor are within the scope of protection of the present application.

[0031] As Figure 1 shown, the embodiment of the present application proposes an image classification method based on hyper-dimensional computation, including the following steps: Step 1, feature extraction of the infrared image is carried out by using a local binary pattern (LBP), so as to obtain a target texture feature.

[0032] Step 2, feature dimension reduction of the target texture feature is carried out by using a principal component analysis (PCA), so as to obtain one-dimensional structural feature information of the target.

[0033] Step 3, the one-dimensional structural feature information of the target is subjected to hyper-dimensional coding by using a high-efficiency multivariate coding method to form a hyper-dimensional vector.

[0034] Step 4, the obtained hyper-dimensional vector is used to construct a hyper-dimensional vector dataset, and the hyper-dimensional vector dataset is divided into a training set, a validation set and a test set; Step 5, a target classification and recognition model based on a binary full connection neural network is constructed. Step 6, the target classification and recognition model based on the binary full connection neural network is trained by using the hyper-dimensional vector dataset, wherein the training set and the validation set are used for model training, and the test set is used for selecting a model with the best performance. Step 7, the hyper-dimensional vector formed after the feature extraction, the feature dimension reduction and the hyper-dimensional coding of the collected infrared image is input into the trained target classification and recognition model for target recognition, so as to obtain a target category.

[0035] The above image classification method based on hyper-dimensional calculation is described in detail as follows: 1, real-time acquisition of an infrared image, feature extraction of the infrared image is carried out by using a local binary pattern, so as to obtain a target texture feature. The local binary pattern is a feature extraction method used for image processing, and is used to describe the local texture feature of an image. Specifically, the local texture feature of an image is extracted by comparing the gray value of a center pixel with that of a field pixel.

[0036] 2, feature dimension reduction of the target texture feature is carried out by using a principal component analysis method. The principal component analysis method is a feature dimension reduction method, which projects high-dimensional data into a low-dimensional space while retaining the variance information of the data to the greatest extent, that is, the key effective information of the feature, so as to extract the most effective feature of the image target.

[0037] 3, the local binary pattern and the principal component analysis method are used to extract the effective feature of the target, so as to solve the problem that a large-size image input is too large to be subjected to hyper-dimensional coding. After the feature extraction and the feature dimension reduction, a feature vector is formed, and the feature dimension of each sample (image) is K dimension, wherein K The value length of the feature vector is [16, 128], and a minimum-maximum value normalization operation is used to normalize the feature vector to the range of [0, 1]. The normalization formula is as follows:

[0038] in, The th feature vector of the current feature vector i 1 eigenvalue, and This represents the minimum and maximum values ​​of the current feature vector. This represents the th normalized feature vector. i Each feature value.

[0039] 4. The normalized feature vectors are transformed into hyperdimensional vectors using an efficient multivariate encoding method. This method improves the speed of hyperdimensional encoding, reduces computational complexity, and minimizes accuracy loss. It involves multiplication and permutation operations in hyperdimensional computation, as well as concepts related to structured density. The overall block diagram of the efficient multivariate encoding method is shown below. Figure 2 As shown. This method first quantizes the feature values ​​of the normalized feature vector (one-dimensional structural feature information of the target). The quantized feature values ​​are then encoded into a hyperdimensional vector using structured density coding. Let the feature vector dimension be 1. G The first feature vector is obtained through structured density encoding. g Numerical hypervectors corresponding to each eigenvalue ,in D For hyperdimensional encoding dimensions, in efficient multivariate encoding methods, it is located at the [number]th [dimension]. g Numerical hypervectors of addresses Permutation operations will be used. G - g This is reflected in actual calculations as a circular right shift. G - g The numerical hypervectors, after undergoing their respective shift operations, are then bound together through multiplication to obtain the final D-dimensional hypervector. Figure 2 In this context, MIN is 0, and MAX is the dimension D of the hyperdimensional encoding. The process of forming this hyperdimensional vector can be expressed by the formula:

[0040] in, R This represents a hyperdimensional vector encoded using an efficient multivariate encoding method. Indicates a circular right shift G -1 position, and so on; This indicates a multiplication operation.

[0041] 5. Eigenvalue quantization refers to the process of quantizing each eigenvalue of an eigenvector into an integer. A specific diagram is shown below. Figure 3As shown, the input feature vector dimension N is 4, and the hyperdimensional vector dimension D is 10000. The actual feature vector dimension is determined based on the time consumption and classification accuracy. This method sets the hyperdimensional vector dimension to 10000 because when the hyperdimensional vector dimension is 10000, the hyperdimensional vectors are essentially orthogonal, resulting in better classification and recognition performance. The formula for feature value quantization is shown below:

[0042] in, D This represents the dimension of a hyperdimensional vector. U F The feature values ​​represent the one-dimensional structural feature information before quantization. U Q Represents the quantized feature values. This indicates rounding operations.

[0043] 6. Structured Density Coding: In the field of random computing, scalars are often represented as random bit vectors, where the encoded scalar determines the probability of generating a 1. Therefore, the method of encoding scalars using the density of 1s in the bit vector is called density-based coding. This invention does not require the scalar encoding representation to be random; therefore, it employs structured density coding, specifically as follows: Figure 4 As shown.

[0044] Taking bipolar coding as an example, where 1 represents 1 and 0 represents -1, the quantized feature values ​​are encoded, with the number of 1s corresponding to the integer value, and other coding bits representing -1. For example... Figure 3 As shown, the quantized integer of the third eigenvalue is 3686. Figure 4 There are 3686 solid circles, used to represent the number of 1s (3686 ones), and hollow circles represent the number of -1s. Note: The binary encoding method is the same, that is, 1 represents 1 and 0 represents 0.

[0045] 7. Binding operation: such as Figure 2 As shown, the binding operation is to perform a multiplication operation on the numerical supervectors after their respective shift operations. The specific calculation is shown in the formula for the process of forming a superdimensional vector mentioned above. For superdimensional vectors using binary encoding, the calculation method of the multiplication operation is the same as that of the XOR operation.

[0046] 8. The hyperdimensional vector dataset refers to the dataset formed into M classes (M1, M2) after LBP feature extraction and PCA feature dimensionality reduction, using an efficient multivariate hyperdimensional encoding method. 1) Hyperdimensional vector dataset. For example, in a hyperdimensional vector dataset, there are 600 samples in each class. The dataset is divided into training set, validation set and test set according to a certain ratio. The training set has 420 samples, the validation set has 90 samples and the test set has 180 samples.

[0047] 9、Binary fully connected neural network is a special type of neural network, which is characterized by limiting the weights and inputs of the neural network to two values (binary values). Binary fully connected neural network is a limit quantization model that uses 0 and 1 or -1 and 1 to represent parameters. The advantage is to meet the input requirement of super vector being all binary, and also to reduce the consumption of storage and computing resources. The method of the present application uses deterministic binarization, XOR operation and straight-through estimator to improve the accuracy of super vector. The overall block diagram is shown in Figure 5 .

[0048] (1) Deterministic binarization When the input value is greater than the set threshold, the output value is +1, otherwise the output value is -1. The mathematical expression of the deterministic binarization operation is as follows:

[0049] Wherein, represents the input floating-point parameter, represents the binary parameter after deterministic binarization. Deterministic binarization has the advantages of simple operation and good effect.

[0050] (2) XOR operation (XOR binary fully connected operation) In binary fully connected neural network, the fully connected layer operation and convolution operation are only multiplication and accumulation operations of +1 and -1, so a new method of using XOR operation instead of traditional fully connected layer operation is derived. Compared with the exclusive or method, one step is less executed on the hardware FPGA, because the hardware only has XOR operation, and the exclusive or operation also needs to perform an inversion operation, which can reduce the consumption of hardware resources and improve the inference speed. The specific operation is shown in Figure 6 . First, the non-binary weight and input feature value are converted to 0 and 1 form by the Sign function (binary sign function); second, the XOR operation is performed to obtain the XOR operation result; finally, the number of features with the XOR operation result of 1 is counted Popcount_xor, and the final result is the total number of features minus twice the Popcount_xor, and the result is -1.

[0051] (3) Straight-through estimator In the binary fully connected neural network, due to the introduction of the deterministic binarization operation, there is a difference between its training process and the training process of the full-precision neural network. Because the derivative of the Sign function (binary sign function) is 0 everywhere (the schematic diagram of the derivative of the Sign function before and after derivation is shown in Figure 7 ), therefore, in the back propagation process, when the gradient reaches the binary activation layer, it becomes 0, which causes the weights of the binary fully connected neural network to be unable to update.

[0052] To solve the problem that the derivative of the Sign function is 0, which leads to the weights of the binary fully connected neural network being unable to update, and to improve the classification accuracy of the hyper-dimensional vector after hyper-dimensional coding, STE (Straight Through Estimate) is used to approximate the derivative of the Sign function. Specifically, the clip function is used to replace the Sign function to realize gradient derivation. The specific calculation method of the clip function is as follows:

[0053] wherein, x is the full-precision weight or activation value in the network, i min ,i max is the set derivable space. The schematic diagram of the clip function before and after derivation is shown in Figure 8 .

[0054] As shown in Figure 5 , in the training stage, the training set and the test set are first imported, and after the weights of the binary fully connected neural network are initialized or updated, the Sign function is used to binarize the weights, the binary linear layer is used for forward propagation, and the XOR operation is used to complete the classification prediction and calculate the loss; in the backward propagation, the clip function with local gradient is used to clip the weights, and then the linear layer and the fully connected layer are used to update the weights, and the optimal binary weights are saved at the end of each training period. In the inference stage, the weights and the test set are directly loaded, and the binarization, XOR operation and Sign function are sequentially executed to output the classification result and evaluate the accuracy, and the whole process does not need to be trained again.

[0055] In the training of the target classification recognition model based on the binary fully connected neural network, the linear layer is used in the forward propagation process to complete the operation with the Sign function, and the clip function with a larger derivable interval is used in the backward propagation process to replace the original Sign function to ensure that the gradient can be calculated. In the network inference process, the Sign function is used to realize the efficient XOR operation of full binarization.

[0056] (4) Binary network architecture Through experiments, the binary fully connected neural network in the embodiment is constructed only by using 2 layers of fully connected layers to reduce the network complexity, and the input layer and the output layer are used, as shown in Figure 9 , specifically, the structure of the target classification recognition model based on the binary fully connected neural network includes an input layer, a first fully connected layer, a first binary activation function layer, a second fully connected layer, a second binary activation function layer and an output layer connected in sequence.

[0057] The mathematical expression of the binary fully connected neural network is as follows:

[0058] wherein, represents an intermediate feature, represents a binary activation function, represents a linear layer. The number of channels of the 2-layer convolution varies by .

[0059] 10、The best performance model refers to: in the model training process, different models are generated by using different hyperparameters, and the optimal model is selected by the test set classification accuracy rate index.

[0060] Based on the same technical concept, the embodiment of the application also provides an electronic device which can realize the process of the image classification method based on hyper-computing provided by the above-mentioned embodiment of the application. In an embodiment, the electronic device can be a server, a terminal device or other electronic device. As shown in the Figure 10 , the electronic device can include: at least one processor, and a memory connected with the at least one processor, and the specific connection medium between the processor and the memory is not limited in the embodiment of the application, Figure 10 for example, the connection between the processor and the memory is through a bus. The bus is represented by a thick line in Figure 10 , and the connection mode between other components is only schematically illustrated and is not limited. The bus can be divided into an address bus, a data bus, a control bus, etc. For convenience, in Figure 10 , only one thick line is used to represent it, but it does not mean that there is only one bus or only one type of bus. Alternatively, the processor can also be called a controller, and the name is not limited.

[0061] In the embodiment of the application, the memory stores instructions executable by the at least one processor, and the at least one processor can execute the foregoing image classification method based on hyper-computing by executing the instructions stored in the memory.

[0062] The processor is the control center of the device, can connect each part of the entire control device through various interfaces and lines, and process data and control the whole device by running or executing the instructions stored in the memory and calling the data stored in the memory. Various functions and processing data of the device, thereby overall monitoring the device.

[0063] In an alternative design, the processor can include one or more processing units, and the processor can integrate an application processor and a modem processor, wherein the application processor mainly processes operating systems, user interfaces, and application programs, and the modem processor mainly processes wireless communications. It can be understood that the above-mentioned modem processor can also not be integrated into the processor. In some embodiments, the processor and the memory can be implemented on the same chip, and in some embodiments, they can also be implemented on separate chips, respectively.

[0064] The processor can be a general purpose processor, such as a CPU, a digital signal processor, an application specific integrated circuit, a field programmable gate array, or other programmable logic device, a discrete gate or transistor logic, a discrete hardware component, and can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general purpose processor can be a microprocessor or any conventional processor. The steps of the image classification method based on hyper-computing disclosed in the embodiments of the present application can be directly embodied as execution by a hardware processor, or executed by a combination of hardware and software modules in the processor.

[0065] The memory is a non-volatile computer readable storage medium, which can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The memory can include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card type memory, random access memory (RAM), static random access memory (SRAM), programmable read only memory (PROM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), magnetic memory, magnetic disk, optical disk, etc. The memory is any other medium capable of carrying or storing desired program code in the form of instructions or data structures and capable of being accessed by a computer, but is not limited to this. The memory in the embodiments of the present application can also be a circuit or any other device capable of realizing the storage function, used for storing program instructions and / or data.

[0066] By designing and programming the processor, the code corresponding to the image classification method based on hyper-computing introduced in the above embodiments can be fixed into the chip, so that the chip can execute the steps of the method of the above embodiments at runtime. How to design and program the processor is a technology known to those skilled in the art, which will not be described here.

[0067] Based on the same inventive concept, the embodiments of the present application also provide a storage medium, which stores computer instructions, when the computer instructions are run on a computer, the computer instructions make the computer execute the image classification method based on hyper-dimensional computation discussed above.

[0068] In some optional embodiments, the various aspects of the image classification method based on hyper-dimensional computation can also be implemented in the form of a program product, which includes program codes, when the program product is run on an apparatus, the program codes are used to make the control device execute the steps of the image classification method based on hyper-dimensional computation according to various exemplary embodiments of the present application described above in the specification.

[0069] It should be noted that although several units or sub-units of the apparatus are mentioned in the foregoing detailed description, such division is merely exemplary and not mandatory. In fact, according to the embodiments of the present application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided into units embodied by multiple units. In addition, although the operations of the method of the present application are described in a particular order in the accompanying drawings, this does not require or imply that the operations must be performed in this particular order, or that all of the illustrated operations must be performed to achieve the desired result. Additionally or alternatively, certain steps can be omitted, multiple steps can be combined into one step, and / or one step can be split into multiple steps.

[0070] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer-usable program code.

[0071] The present application is described with reference to the flowcharts and / or block diagrams of the methods, apparatus (systems) and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as a combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one flow or multiple flows and / or blocksFigure 1 means for performing the function specified by the block or blocks.

[0072] programmable data processing apparatus, to produce a machine, such that the instructions which execute on the computer or other programmable data processing apparatus

[0073] In this document, the terms "computer-readable medium" and "computer-readable storage medium" are used to generally refer to media such as removable storage units or memory, including both volatile and nonvolatile memory modules. Such

[0074] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 flow or flows and / or blocks Figure 1 means for performing the function specified by the block or blocks.

[0075] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the Figure 1 flow or flows and / or blocks Figure 1 steps of a function specified by the block or blocks.

[0076] The above description is only preferred embodiments of the present application, and is not intended to limit the present application. The present application can have various modifications and changes, and any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. An image classification method based on hyperdimensional computation, characterized in that, include: The target texture features are obtained by extracting features from infrared images using local binary mode. Principal component analysis is used to reduce the dimensionality of the target texture features to obtain one-dimensional structural feature information of the target. The one-dimensional structural feature information of the target is encoded into a hyperdimensional vector using an efficient multivariate encoding method. Construct a hyperdimensional vector dataset from the obtained hyperdimensional vectors; Construct a target classification and recognition model based on a binary fully connected neural network; A target classification and recognition model based on a binary fully connected neural network was trained using a hyperdimensional vector dataset. The acquired infrared images are processed through feature extraction, feature dimensionality reduction, and hyperdimensional encoding to form a hyperdimensional vector, which is then fed into a trained target classification and recognition model for target recognition to obtain the target category.

2. The image classification method based on hyperdimensional computation according to claim 1, characterized in that, The one-dimensional structural feature information of the target is encoded into a hyperdimensional vector using an efficient multivariate encoding method, including: The feature values ​​of the one-dimensional structural feature information of the target are quantized, and the quantized feature values ​​are formed into a hyperdimensional vector through structured density encoding. Let the dimension of the feature vector be 1. G The first feature vector is obtained through structured density encoding. g The hyperdimensional vector corresponding to each eigenvalue ,in D Encoding dimensions for hyperdimensionality; For the position located at g A hyperdimensional vector of addresses Using permutation operations G - g This is reflected in actual calculations as a circular right shift. G - g Bit; After undergoing their respective shift operations, the numerical hypervectors are then bound together through multiplication to obtain the final D-dimensional hypervector, represented as: ;in, R This represents a hyperdimensional vector encoded using an efficient multivariate encoding method. Indicates a circular right shift G -1 position, and so on; This indicates a multiplication operation.

3. The image classification method based on hyperdimensional computation according to claim 2, characterized in that, The formula for eigenvalue quantization is shown below: in, D This represents the dimension of a hyperdimensional vector. U F The feature values ​​represent the one-dimensional structural feature information before quantization. U Q Represents the quantized feature values. This indicates rounding operations.

4. The image classification method based on hyperdimensional computation according to claim 2, characterized in that, The structured density coding employs either bipolar coding or binary coding.

5. The image classification method based on hyperdimensional computation according to claim 1, characterized in that, The target classification and recognition model based on a binary fully connected neural network comprises an input layer, a first fully connected layer, a first binary activation function layer, a second fully connected layer, a second binary activation function layer, and an output layer connected in sequence.

6. The image classification method based on hyperdimensional computation according to claim 5, characterized in that, The calculation process of the fully connected layer in the target classification and recognition model based on a binary fully connected neural network includes: Perform deterministic binarization: Convert non-binary weights and input feature values ​​into 0 and 1 forms using the Sign function; Perform an XOR operation to obtain the XOR result; Popcount_xor is the number of features whose XOR operation result is 1. The final result is the total number of features minus twice the Popcount_xor.

7. The image classification method based on hyperdimensional computation according to claim 1, characterized in that, When training a target classification and recognition model based on a binary fully connected neural network, a linear layer is used in conjunction with the Sign function during the forward propagation process, while the Clip function replaces the Sign function during the back propagation process to ensure that the gradient is computable; during network inference, the Sign function is used to implement the XOR operation.

8. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, which executes the instructions stored in the memory to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store instructions that, when executed, cause the method as described in any one of claims 1-7 to be implemented.

10. A computer program product, characterized in that, When the computer program product is invoked by a computer, it causes the computer to perform the method as described in any one of claims 1-7.

Images