CN110929697B - Neural network target identification method and system based on residual error structure - Google Patents

Neural network target identification method and system based on residual error structure Download PDF

Info

Publication number
CN110929697B
CN110929697B CN201911301003.8A CN201911301003A CN110929697B CN 110929697 B CN110929697 B CN 110929697B CN 201911301003 A CN201911301003 A CN 201911301003A CN 110929697 B CN110929697 B CN 110929697B
Authority
CN
China
Prior art keywords
layer
feature
convolution
neural network
training data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201911301003.8A
Other languages
Chinese (zh)
Other versions
CN110929697A (en
Inventor
但波
付哲泉
杨富程
戢治洪
高山
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Naval Aeronautical University
Original Assignee
Naval Aeronautical University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Naval Aeronautical University filed Critical Naval Aeronautical University
Priority to CN201911301003.8A priority Critical patent/CN110929697B/en
Publication of CN110929697A publication Critical patent/CN110929697A/en
Application granted granted Critical
Publication of CN110929697B publication Critical patent/CN110929697B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/07—Target detection

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to a neural network target identification method and a system based on a residual error structure, wherein the method comprises the following steps: acquiring target image data, and labeling the target image data according to a target category to form training data; carrying out normalization and zero equalization on the training data to obtain processed training data; constructing a neural network of a residual structure, wherein the neural network of the residual structure comprises a convolution module layer, a first full-connection layer and an output layer which are sequentially connected, and training the neural network of the residual structure by using a joint loss function according to the processed training data to obtain a trained network model; acquiring target image data to be recognized, and performing normalization and zero-averaging on the target image data to be recognized to obtain processed data to be recognized; and inputting the processed data to be recognized into the trained network model, and outputting a recognition result. The method can improve the accuracy of target identification.

Description

Neural network target identification method and system based on residual error structure
Technical Field
The invention relates to the field of target identification, in particular to a neural network target identification method and system based on a residual error structure.
Background
Designing a good neural network structure is one of the efficient and challenging methods to improve classification performance. On the premise of sufficient sample quantity of the data set, the learning capacity of the model can be improved by increasing the depth and the width of the neural network, and both AlexNet and VGG structures prove that the identification accuracy of the model is positively correlated with the depth of the network within a certain range. However, as the depth of the network increases, the convolutional neural network may have the problems of gradient explosion, disappearance and saturation of the network recognition rate in the back propagation process during training. I.e., the network cannot learn new valid features, and recognition accuracy decreases as network depth increases. To address this problem, hoxamine et al propose a residual structure that preserves the integrity of the shallow features through hopping connections, avoiding network performance degradation with increasing depth. But the residual structure network still needs to increase the network depth to improve the recognition effect.
Disclosure of Invention
The invention aims to provide a neural network target identification method and system based on a residual error structure, which can improve the accuracy of target identification.
In order to achieve the purpose, the invention provides the following scheme:
a neural network target identification method based on a residual error structure, the target identification method comprises the following steps:
acquiring target image data, and labeling the target image data according to a target category to form training data;
normalizing and zero-averaging the training data to obtain processed training data;
constructing a neural network of a residual structure, the neural network of the residual structure comprising: the system comprises a convolution module layer, a first full-connection layer and an output layer which are sequentially connected, wherein the convolution module layer comprises a plurality of convolution modules, the convolution modules perform feature extraction and fusion on different layers of the processed training data to obtain fusion features, the first full-connection layer sorts the fusion features according to a set sequence to obtain one-dimensional feature vectors, and the output layer classifies the training data according to the one-dimensional feature vectors by using a classifier and outputs classification results;
training the neural network with the constructed residual error structure by using a joint loss function according to the processed training data to obtain a trained network model;
acquiring target image data to be recognized, and carrying out normalization and zero-averaging on the target image data to be recognized to obtain processed data to be recognized;
and inputting the processed data to be recognized into the trained network model, and outputting a recognition result.
Optionally, the convolution module performs feature extraction and fusion on different layers of the processed training data to obtain a fusion feature, which specifically includes:
performing a first layer of convolution operation on the processed training data to obtain a first layer of characteristics, wherein the step length of the first layer of convolution is S1The size of the convolution kernel is K1×1；
Equally dividing the first layer of characteristics into Z parts, and performing convolution operation on each corresponding characteristic by adopting Z second layer convolutions to obtain Z second layer characteristics, wherein the step length of the second layer convolution is S2The size of the convolution kernel is K2×1；
Combining the Z second layer characteristics to obtain combined second layer characteristics, and performing convolution operation on the combined second layer characteristics by adopting a third layer of convolution to obtain a third layer of characteristics, wherein the step length of the third layer of convolution is S1The size of the convolution kernel is K1×1；
Dividing the third layer of characteristics into two parts, and marking the three parts as a first characteristic and a second characteristic;
performing a fourth layer of convolution operation on the processed training data to obtain a fourth layer of characteristics, wherein the step length of the fourth layer of convolution is S2The size of the convolution kernel is K1×1；
Dividing the fourth layer of features into two parts, and marking the two parts as third features and fourth features, wherein the dimensions of the first features and the three features are the same, and the dimensions of the second features and the fourth features are the same;
adding each element in the first feature and each element in the third feature to obtain an added first feature;
combining the second feature and the fourth feature to obtain a combined second feature;
and combining the added first feature and the combined second feature to obtain a fused feature.
Optionally, the neural network of the residual error structure further includes: and the output of the initial convolutional layer is connected with the input of the convolutional module layer, and the initial convolutional layer extracts the characteristics of the processed training data to obtain initial characteristics.
Optionally, the neural network of the residual error structure further includes: a second full link layer, the output of the second full link layer and the input of the output layer are connected, the second full link layer is right the first fusion feature is visualized.
Optionally, S is11, said S 22, said K 11, said K2＝2。
A neural network target recognition system based on a residual structure, the target recognition system comprising:
the training data acquisition module is used for acquiring target image data and marking the target image data according to target categories to form training data;
the processed training data determining module is used for carrying out normalization and zero equalization on the training data to obtain processed training data;
a neural network construction module of a residual structure, configured to construct a neural network of a residual structure, where the neural network of a residual structure includes: the system comprises a convolution module layer, a first full-connection layer and an output layer which are sequentially connected, wherein the convolution module layer comprises a plurality of convolution modules, the convolution modules perform feature extraction and fusion on different layers of the processed training data to obtain fusion features, the first full-connection layer sorts the fusion features according to a set sequence to obtain one-dimensional feature vectors, and the output layer classifies the training data according to the one-dimensional feature vectors by using a classifier and outputs classification results;
the network model training module is used for training the neural network with the constructed residual error structure by using a joint loss function according to the processed training data to obtain a trained network model;
the processed data to be identified acquisition module is used for acquiring target image data to be identified, and carrying out normalization and zero-averaging on the target image data to be identified to obtain processed data to be identified;
and the recognition module is used for inputting the processed data to be recognized into the trained network model and outputting a recognition result.
Optionally, the convolution module performs feature extraction and fusion on different layers of the processed training data to obtain a fusion feature, which specifically includes:
a first layer feature determination unit, configured to perform a first layer convolution operation on the processed training data to obtain a first layer feature, where a step length of the first layer convolution is S1The size of the convolution kernel is K1×1；
A second layer feature determining unit, configured to divide the first layer features into Z parts equally, and perform convolution operation on each corresponding feature by using Z second layer convolutions to obtain Z second layer features, where a step length of the second layer convolution is S2The size of the convolution kernel is K2×1；
A third layer feature determining unit, configured to combine the Z second layer features to obtain a combined second layer feature, and perform convolution operation on the combined second layer feature by using a third layer convolution to obtain a third layer feature, where a step length of the third layer convolution is S1The size of the convolution kernel is K1×1；
A first feature and second feature determining unit, configured to divide the third layer feature into two parts, which are denoted as a first feature and a second feature;
a fourth layer feature determining unit, configured to perform a fourth layer convolution operation on the processed training data to obtain a fourth layer feature, where a step length of the fourth layer convolution is S2Size of convolution kernelIs K1×1；
A third feature and fourth feature determining unit, configured to divide the fourth layer feature into two parts, which are denoted as a third feature and a fourth feature, where the first feature and the third feature have the same dimension, and the second feature and the fourth feature have the same dimension;
a first feature adding unit configured to add elements in the first feature and elements in the third feature to obtain an added first feature;
a second feature merging unit, configured to merge the second feature and the fourth feature to obtain a merged second feature;
and a fusion feature determining unit, configured to combine the added first feature and the combined second feature to obtain a fusion feature.
Optionally, the neural network of the residual error structure further includes: and the output of the initial convolutional layer is connected with the input of the convolutional module layer, and the initial convolutional layer is used for extracting the characteristics of the processed training data to obtain initial characteristics.
Optionally, the neural network of the residual error structure further includes: a second fully-connected layer, the output of which is connected to the input of the output layer, the second fully-connected layer being used to visualize the first fused feature.
Optionally, S is11, said S 22, said K 11, said K2＝2。
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
the invention provides a neural network target identification method and a system based on a residual error structure, wherein the method comprises the following steps: acquiring target image data, and labeling the target image data according to a target category to form training data; carrying out normalization and zero equalization on the training data to obtain processed training data; constructing a neural network of a residual structure, wherein the neural network of the residual structure comprises a convolution module layer, a first full-connection layer and an output layer which are sequentially connected, and training the neural network of the residual structure by using a joint loss function according to the processed training data to obtain a trained network model; acquiring target image data to be recognized, and performing normalization and zero-averaging on the target image data to be recognized to obtain processed data to be recognized; and inputting the processed data to be recognized into the trained network model, and outputting a recognition result. The method can improve the accuracy of target identification.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.
FIG. 1 is a flowchart of a neural network target identification method based on a residual error structure according to an embodiment of the present invention;
fig. 2 is a schematic structural diagram of a CNN according to an embodiment of the present invention;
FIG. 3 is a diagram illustrating a neural network structure of a residual structure according to an embodiment of the present invention;
FIG. 4 is a diagram illustrating a conventional residual structure;
FIG. 5 is a diagram illustrating a convolution module according to an embodiment of the present invention;
fig. 6 is a schematic structural diagram of a neural network target recognition system based on a residual error structure according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention aims to provide a neural network target identification method and system based on a residual error structure, which can improve the accuracy of target identification.
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.
Fig. 1 is a flowchart of a neural network target identification method based on a residual error structure according to an embodiment of the present invention, and as shown in fig. 1, the neural network target identification method based on the residual error structure includes:
s1: and acquiring target image data, and labeling the target image data according to a target category to form training data.
S2: and carrying out normalization and zero equalization on the training data to obtain processed training data.
S3: constructing a neural network of a residual structure, the neural network of the residual structure comprising: the system comprises a convolution module layer, a first full-connection layer and an output layer which are sequentially connected, wherein the convolution module layer comprises a plurality of convolution modules, the convolution modules extract and fuse features of different layers of processed training data to obtain fusion features, the first full-connection layer sorts the fusion features according to a set sequence to obtain one-dimensional feature vectors, and the output layer classifies the training data according to the one-dimensional feature vectors by using a classifier and outputs classification results.
Specifically, the neural network of the residual error structure further includes: and the output of the initial convolutional layer is connected with the input of the convolutional module layer, and the initial convolutional layer extracts the characteristics of the processed training data to obtain initial characteristics.
The neural network of the residual structure further comprises: a second full link layer, the output of the second full link layer and the input of the output layer are connected, the second full link layer is right the first fusion feature is visualized.
The convolution module extracts and fuses features of different layers of the processed training data to obtain fusion features, and the method specifically comprises the following steps:
performing a first layer of convolution operation on the processed training data to obtain a first layer of characteristics, wherein the step length of the first layer of convolution is S1The size of the convolution kernel is K1×1；
Equally dividing the first layer of characteristics into Z parts, and performing convolution operation on each corresponding characteristic by adopting Z second layer convolutions to obtain Z second layer characteristics, wherein the step length of the second layer convolution is S2The size of the convolution kernel is K2×1；
Combining the Z second layer characteristics to obtain combined second layer characteristics, and performing convolution operation on the combined second layer characteristics by adopting a third layer of convolution to obtain a third layer of characteristics, wherein the step length of the third layer of convolution is S1The size of the convolution kernel is K1×1；
Dividing the third layer of characteristics into two parts, and marking the three parts as a first characteristic and a second characteristic;
performing a fourth layer of convolution operation on the processed training data to obtain a fourth layer of characteristics, wherein the step length of the fourth layer of convolution is S2The size of the convolution kernel is K1×1；
Dividing the fourth layer of features into two parts, and marking the two parts as third features and fourth features, wherein the dimensions of the first features and the three features are the same, and the dimensions of the second features and the fourth features are the same;
adding each element in the first feature and each element in the third feature to obtain an added first feature;
combining the second feature and the fourth feature to obtain a combined second feature;
and combining the added first feature and the combined second feature to obtain a fused feature.
Said S 11, said S 22, said K 11, said K2＝2。
S4: and training the neural network with the constructed residual error structure by using a joint loss function according to the processed training data to obtain a trained network model.
S5: acquiring target image data to be recognized, and performing normalization and zero-averaging on the target image data to be recognized to obtain processed data to be recognized.
S6: and inputting the processed data to be recognized into the trained network model, and outputting a recognition result.
The following is a detailed discussion of the steps:
the convolutional neural network is a feed-forward neural network containing convolutional calculation, and has the characteristic of translation invariance in the calculation process, so that complex preprocessing such as HRRP (High Range Resolution Profile) data alignment can be avoided, the convolutional neural network has better robustness, and a High-Resolution Range image (HRRP) of a radar target represents the projection of a target scattering center in the radar-target direction and contains rich target characteristics.
The model used by the invention is based on a convolutional neural network, and the basic structure of the model is introduced firstly, and mainly comprises the following components: input layer, convolution layer, pooling layer, full connection layer, output layer 5 parts.
The CNN structure for HRRP is shown in fig. 2, and fig. 2 shows the classification process of 10 kinds of targets for HRRP data with a length of 128 by CNN.
The input layer is the starting point of the neural network, and generally requires simple preprocessing of the data so that the data have the same dimensionality and satisfy the same distribution characteristics. The preprocessing can reduce the influence of amplitude disturbance on network extraction characteristics of different HRRP data of the same target, improve the robustness of the model, and facilitate finding the minimum value more directly in the iteration process of the gradient descent method, so that the model is converged more quickly.
The data are normalized and zero-averaged in both S2 and S5:
(1) and (6) normalizing. And carrying out normalization processing on the data amplitude. For the nth HRRP data x (n), the amplitude normalized data is represented as x'n＝xn/max(|xn|). Where max (| x)n|) is the maximum of the absolute values of all elements in the HRRP data.
(2) And (4) zero mean value. And respectively subtracting the average value of the HRRP data from each element of the normalized HRRP data.
The convolutional layer mainly functions to extract features of input data, and generally, the convolutional layer may include a plurality of convolutional kernels inside. In fig. 2, the first convolutional layer contains 16 convolutional kernels, and the second convolutional layer contains 32 convolutional kernels, and each convolutional kernel element corresponds to a weight coefficient and an offset value. Performing convolution calculation on input data by a convolution kernel and adding an offset value, and then activating by using an activation function, wherein the output is the extracted feature, and the calculation process can be represented by a formula (1):
Figure BDA0002321774260000081
wherein x represents a feature item; m represents a set of input feature items; k is a convolution kernel; b is a deviation value; l is the layer number of the network structure; i is the convolution kernel number; j is the serial number of the feature item channel; f is the activation function. The parameters of the convolutional layer comprise the size of a convolutional kernel, step length, filling type and an activation function, common activation functions comprise a Sigmoid function, a Relu function and the like, and different parameters endow the convolutional layer with different characteristics.
The pooling layer mainly has the functions of selecting and reducing the extracted features of the convolutional layer through down-sampling, and the commonly used pooling layer has maximum pooling, mean pooling, mixed pooling and the like.
The fully-connected layer is usually placed at the rear part of the neural network, and mainly has the functions of arranging the features obtained at the previous layer in sequence to obtain a one-dimensional vector, outputting the result of obtaining the whole CNN through an output layer classifier, wherein the commonly used classifier comprises softmax, a support vector machine and the like. In the object recognition problem, the output of the CNN may be the class, size, center coordinates, and the like of the object. The CNN learning process for supervised learning generally performs iterative updating of parameters by back propagation, and obtains a stable recognition result by minimizing an error value calculated by a loss function.
Fig. 3 is a schematic diagram of a neural network structure of a residual structure according to an embodiment of the present invention, and as shown in fig. 3, the neural network of the residual structure includes an initial convolution layer, a plurality of convolution modules of the same topology structure connected in sequence, and two full-connection layers, where the dimension of the second full-connection layer is 2, so as to facilitate visualization of features extracted by a model and analysis of a clustering effect of the features.
The input is a High Range Resolution Profile (HRRP) of the radar target, which represents the projection of the target scattering center in the radar-target direction, and contains abundant target features. The input is a one-dimensional vector, preferably an even number, and the dimension of the input one-dimensional vector should be n times of 2, because the output of the convolution module is halved in characteristic dimension and doubled in layer number relative to the input, to ensure that the data dimension can be always divided by 2, the number in the bracket represents the data dimension of the HRRP sample after passing through the layer.
And determining the output data dimensionality of the plurality of continuous convolution modules and the first full-connection layer according to the number of the convolution modules. The result of the final output layer is one-dimensional data corresponding to the target class, here the number of target classes 13.
The initial convolutional layer is used for reducing dimensions, a one-dimensional convolution kernel with the dimension of 7 x 1 is selected for the initial convolutional layer, a convolution kernel with a relatively large dimension is selected for the first layer of the network, so that the extraction of the corresponding features such as contours and textures in the target HRRP data is facilitated, and batch normalization and Relu activation operation are performed on the extracted features after each convolution operation in the model.
The depth of the neural network is important, and the deep convolutional neural network can extract and fuse features of different layers for end-to-end target identification. However, the network layer number is deepened, which causes the problem of saturated recognition accuracy, and the residual structure is generally introduced to overcome the problem, and the structure is shown in fig. 4.
The residual block in the residual structure is composed of convolutional layers, the number of convolutional layers in fig. 4 is 2, the output of the residual structure is the sum of the input features and the output of the last convolutional layer, and is represented by formula (2):
xl+1＝F(xl)+xl (2)
wherein x isl、xl+1Respectively representing the input and output characteristics of the l-th layer residual structure, F (x)l) Representing a mapping of the residual block.
Studies have shown that F (x) is mapped by fittingl) Instead of the mapping F (x) of the desired fitl)+xlThe problem of saturated accuracy of deep network identification can be effectively relieved. In extreme cases, if the network extracts the optimal features required for classification, the residual structure only needs to perform equal mapping of jump connection to ensure the highest identification accuracy, and for the neural network, the zero setting of the residual block is more efficient than the fitting of the equal mapping by using the multilayer neural network.
According to the convolution module structure obtained by improving the residual structure, as shown in fig. 5, mx 1 × N in the graph represents one-dimensional data with mx 1 input characteristics, the number of characteristic layers is N, s is the moving step length of a convolution kernel, the unlabeled step length defaults to 1, the convolution module is set to be a highly modular network structure, and the expandability is high.
The characteristics extracted by the upper network are used as the input of the network of the upper network, and the input passes through 2 branches. In the left branch, firstly, features between different layers are fused by using a 1 × 1 first layer convolution kernel, the fused features are equally divided into a plurality of branches (represented by x in the figure) in the layer number, each branch has 3 layers of features, each branch is respectively subjected to feature extraction by using a 3 × 1 second layer convolution kernel, the step length is 2, the feature layer number is unchanged, and the dimension is halved. And then splicing the characteristics of the plurality of branches, determining the size of x according to the complexity of a classification task, uniformly selecting a 3 multiplied by 1 small-scale second layer convolution kernel to reduce the structural design difficulty and ensure the identification effect, performing characteristic fusion on the spliced characteristics by using a 1 multiplied by 1 third layer convolution kernel again and increasing the number of characteristic layers, and dividing the characteristics into two parts according to the layers to prepare for the characteristic fusion of the subsequent two branches. The right branch directly uses a convolution kernel of 1 × 1 to perform feature fusion on the input and increase the number of feature layers, and at the same time, divides the features into two parts according to the number of layers, and performs addition and splicing operations with the features corresponding to the left branch, as shown in fig. 5.
The output of the convolution module is reduced by half in characteristic dimension and doubled in layer number relative to the input. The effect of the right branch is similar to that of a residual error network, each layer of the network module can acquire information from a loss function and an original input signal, characteristics and gradients can be more effectively transmitted, the utilization rate of shallow characteristics is improved, and the problems of gradient loss and identification rate saturation which are possibly generated along with continuous deepening of the network are solved.
The first full-connection layer is mainly used for arranging the features obtained from the previous layer in sequence to obtain a one-dimensional vector.
The penalty function is used to measure the difference between the predicted value and the true value, and is generally denoted by L (y _, y), where y _ denotes the predicted value and y denotes the true value, and for multi-class convolutional neural networks, Softmax Loss (SL) is typically used as the penalty function. However, from a clustering perspective, SL extracted features may appear where the intra-class distance is greater than the inter-class distance. Meanwhile, the features extracted by the SL are divergent during visualization, and when the target categories are too many, feature overlapping is caused, which is not beneficial to target classification.
When classifying objects, it is not only desirable that features be separable, but also that there be a large difference between features. The CenterLoss (CL) loss function can construct a class center for each class of target features, the distance in the class is reduced by punishing the features far away from the class center, and the parameter lambda is introduced to fuse CL and SL to obtain a joint loss function L, as shown in a formula (3):
Figure BDA0002321774260000101
wherein L issRepresents SL; l iscRepresents CL; λ represents L in LcThe occupied weight needs to be reasonably set; x is the number ofi∈dRepresenting the ith depth feature, and d representing the feature dimension;
Figure BDA0002321774260000102
representing weight matrix
Figure BDA0002321774260000103
Column j of (1); b is an element ofnRepresenting a bias term; m and n respectively represent the number of data in each batch and the number of types of targets during batch processing;
Figure BDA0002321774260000104
denotes the y thiThe center of the features of each of the categories,
Figure BDA0002321774260000105
continuously updated as the characteristics of each batch of data change.
L in the joint loss functioncRelative to xiGradient of and
Figure BDA0002321774260000106
the update equation of (a) is as follows:
Figure BDA0002321774260000107
Figure BDA0002321774260000111
wherein when yiFor class j targets, the identification is correct, where δ (-) equals 1, otherwise δ (-) equals 0.
The present invention also provides a neural network target recognition system based on a residual error structure, as shown in fig. 6, the target recognition system includes:
a training data obtaining module 201, configured to obtain target image data, and label the target image data according to a target category to form training data;
a processed training data determining module 202, configured to perform normalization and zero-averaging on the training data to obtain processed training data;
a residual structured neural network constructing module 203, configured to construct a residual structured neural network, where the residual structured neural network includes: the system comprises a convolution module layer, a first full-connection layer and an output layer which are sequentially connected, wherein the convolution module layer comprises a plurality of convolution modules, the convolution modules perform feature extraction and fusion on different layers of the processed training data to obtain fusion features, the first full-connection layer sorts the fusion features according to a set sequence to obtain one-dimensional feature vectors, and the output layer classifies the training data according to the one-dimensional feature vectors by using a classifier and outputs classification results;
a network model training module 204, configured to train the neural network with the constructed residual structure by using a joint loss function according to the processed training data, so as to obtain a trained network model;
the processed data to be recognized acquiring module 205 is configured to acquire target image data to be recognized, and perform normalization and zero-averaging on the target image data to be recognized to obtain processed data to be recognized;
and the recognition module 206 is configured to input the processed data to be recognized into the trained network model, and output a recognition result.
Preferably, the convolution module performs feature extraction and fusion on different layers of the processed training data to obtain a fusion feature, and specifically includes:
a first layer feature determination unit, configured to perform a first layer convolution operation on the processed training data to obtain a first layer feature, where a step length of the first layer convolution is S1The size of the convolution kernel is K1×1；
A second layer feature determining unit, configured to divide the first layer features into Z parts equally, and perform convolution operation on each corresponding feature by using Z second layer convolutions to obtain Z second layer features, where a step length of the second layer convolution is S2The size of the convolution kernel is K2×1；
A third layer feature determining unit, configured to combine the Z second layer features to obtain a combined second layer feature, and perform convolution operation on the combined second layer feature by using a third layer convolution to obtain a third layer feature, where a step length of the third layer convolution is S1The size of the convolution kernel is K1×1；
A first feature and second feature determining unit, configured to divide the third layer feature into two parts, which are denoted as a first feature and a second feature;
a fourth layer feature determining unit, configured to perform a fourth layer convolution operation on the processed training data to obtain a fourth layer feature, where a step length of the fourth layer convolution is S2The size of the convolution kernel is K1×1；
A third feature and fourth feature determining unit, configured to divide the fourth layer feature into two parts, which are denoted as a third feature and a fourth feature, where the first feature and the third feature have the same dimension, and the second feature and the fourth feature have the same dimension;
a first feature adding unit configured to add elements in the first feature and elements in the third feature to obtain an added first feature;
a second feature merging unit, configured to merge the second feature and the fourth feature to obtain a merged second feature;
and a fusion feature determining unit, configured to combine the added first feature and the combined second feature to obtain a fusion feature.
Preferably, the neural network of the residual structure further comprises: and the output of the initial convolutional layer is connected with the input of the convolutional module layer, and the initial convolutional layer is used for extracting the characteristics of the processed training data to obtain initial characteristics.
Preferably, the neural network of the residual structure further comprises: a second fully-connected layer, the output of which is connected to the input of the output layer, the second fully-connected layer being used to visualize the first fused feature.
Preferably, said S is11, said S 22, said K 11, said K2＝2。
The invention provides a high-efficiency extensible neural network with an improved residual error structure. The recognition effect of the deep network is achieved under the condition of reducing network parameters; meanwhile, through the design of a modular structure, the method provided by the invention can be efficiently expanded to adapt to classification tasks with different difficulties, and compared with the traditional algorithm, the method can obtain higher identification accuracy.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system disclosed by the embodiment, the description is relatively simple because the system corresponds to the method disclosed by the embodiment, and the relevant points can be referred to the method part for description.
The principles and embodiments of the present invention have been described herein using specific examples, which are provided only to help understand the method and the core concept of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the above, the present disclosure should not be construed as limiting the invention.

Claims (8)

1. A neural network target identification method based on a residual error structure is characterized by comprising the following steps:
acquiring target image data, and labeling the target image data according to a target category to form training data;
normalizing and zero-averaging the training data to obtain processed training data;
constructing a neural network of a residual structure, the neural network of the residual structure comprising: the system comprises a convolution module layer, a first full-connection layer and an output layer which are sequentially connected, wherein the convolution module layer comprises a plurality of convolution modules, the convolution modules perform feature extraction and fusion on different layers of the processed training data to obtain fusion features, the first full-connection layer sorts the fusion features according to a set sequence to obtain one-dimensional feature vectors, and the output layer classifies the training data according to the one-dimensional feature vectors by using a classifier and outputs classification results;
training the neural network with the constructed residual error structure by using a joint loss function according to the processed training data to obtain a trained network model;
acquiring target image data to be recognized, and carrying out normalization and zero-averaging on the target image data to be recognized to obtain processed data to be recognized;
inputting the processed data to be recognized into the trained network model, and outputting a recognition result;
the convolution module extracts and fuses features of different layers of the processed training data to obtain fusion features, and the method specifically comprises the following steps:
performing a first layer of convolution operation on the processed training data to obtain a first layer of characteristics, wherein the step length of the first layer of convolution is S1The size of the convolution kernel is K1×1；
Equally dividing the first layer of characteristics into Z parts, and performing convolution operation on each corresponding characteristic by adopting Z second layer convolutions to obtain Z second layer characteristics, wherein the step length of the second layer convolution is S2The size of the convolution kernel is K2×1；
Combining the Z second layer characteristics to obtain combined second layer characteristics, and performing convolution operation on the combined second layer characteristics by adopting a third layer of convolution to obtain a third layer of characteristics, wherein the step length of the third layer of convolution is S1The size of the convolution kernel is K1×1；
Dividing the third layer of characteristics into two parts, and marking the three parts as a first characteristic and a second characteristic;
performing a fourth layer of convolution operation on the processed training data to obtain a fourth layer of characteristics, wherein the step length of the fourth layer of convolution is S2The size of the convolution kernel is K1×1；
Dividing the fourth layer of features into two parts, and marking the two parts as third features and fourth features, wherein the dimensions of the first features and the third features are the same, and the dimensions of the second features and the fourth features are the same;
adding each element in the first feature and each element in the third feature to obtain an added first feature;
combining the second feature and the fourth feature to obtain a combined second feature;
and combining the added first feature and the combined second feature to obtain a fused feature.
2. The method of claim 1, wherein the neural network based on the residual structure further comprises: and the output of the initial convolutional layer is connected with the input of the convolutional module layer, and the initial convolutional layer extracts the characteristics of the processed training data to obtain initial characteristics.
3. The method of claim 1, wherein the neural network based on the residual structure further comprises: a second full link layer, the output of the second full link layer and the input of the output layer are connected, the second full link layer is right the fusion feature is visualized.
4. The method of claim 1, wherein the S is a function of a number of parameters of the neural network11, said S22, said K11, said K2＝2。
5. A neural network target recognition system based on a residual structure, the target recognition system comprising:
the training data acquisition module is used for acquiring target image data and marking the target image data according to target categories to form training data;
the processed training data determining module is used for carrying out normalization and zero equalization on the training data to obtain processed training data;
a neural network construction module of a residual structure, configured to construct a neural network of a residual structure, where the neural network of a residual structure includes: the system comprises a convolution module layer, a first full-connection layer and an output layer which are sequentially connected, wherein the convolution module layer comprises a plurality of convolution modules, the convolution modules perform feature extraction and fusion on different layers of the processed training data to obtain fusion features, the first full-connection layer sorts the fusion features according to a set sequence to obtain one-dimensional feature vectors, and the output layer classifies the training data according to the one-dimensional feature vectors by using a classifier and outputs classification results;
the network model training module is used for training the neural network with the constructed residual error structure by using a joint loss function according to the processed training data to obtain a trained network model;
the processed data to be identified acquisition module is used for acquiring target image data to be identified, and carrying out normalization and zero-averaging on the target image data to be identified to obtain processed data to be identified;
the recognition module is used for inputting the processed data to be recognized into the trained network model and outputting a recognition result;
the convolution module extracts and fuses features of different layers of the processed training data to obtain fusion features, and the method specifically comprises the following steps:
a first layer feature determination unit, configured to perform a first layer convolution operation on the processed training data to obtain a first layer feature, where a step length of the first layer convolution is S1The size of the convolution kernel is K1×1；
A second layer feature determining unit, configured to divide the first layer features into Z parts equally, and perform convolution operation on each corresponding feature by using Z second layer convolutions to obtain Z second layer features, where a step length of the second layer convolution is S2The size of the convolution kernel is K2×1；
A third layer feature determining unit, configured to combine the Z second layer features to obtain a combined second layer feature, and perform convolution operation on the combined second layer feature by using a third layer convolution to obtain a third layer feature, where a step length of the third layer convolution is S1The size of the convolution kernel is K1×1；
A first feature and second feature determining unit, configured to divide the third layer feature into two parts, which are denoted as a first feature and a second feature;
a fourth layer feature determining unit, configured to perform a fourth layer convolution operation on the processed training data to obtain a fourth layer feature, where a step length of the fourth layer convolution is S2The size of the convolution kernel is K1×1；
A third feature and fourth feature determining unit, configured to divide the fourth layer feature into two parts, which are denoted as a third feature and a fourth feature, where the first feature and the third feature have the same dimension, and the second feature and the fourth feature have the same dimension;
a first feature adding unit configured to add elements in the first feature and elements in the third feature to obtain an added first feature;
a second feature merging unit, configured to merge the second feature and the fourth feature to obtain a merged second feature;
and a fusion feature determining unit, configured to combine the added first feature and the combined second feature to obtain a fusion feature.
6. The residual structure-based neural network target recognition system of claim 5, wherein the residual structure-based neural network further comprises: and the output of the initial convolutional layer is connected with the input of the convolutional module layer, and the initial convolutional layer is used for extracting the characteristics of the processed training data to obtain initial characteristics.
7. The residual structure-based neural network target recognition system of claim 5, wherein the residual structure-based neural network further comprises: and the output of the second full connection layer is connected with the input of the output layer, and the second full connection layer is used for visualizing the fusion features.
8. According to claimThe neural network target recognition system based on the residual error structure, wherein S is11, said S22, said K11, said K2＝2。
CN201911301003.8A 2019-12-17 2019-12-17 Neural network target identification method and system based on residual error structure Active CN110929697B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911301003.8A CN110929697B (en) 2019-12-17 2019-12-17 Neural network target identification method and system based on residual error structure

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911301003.8A CN110929697B (en) 2019-12-17 2019-12-17 Neural network target identification method and system based on residual error structure

Publications (2)

Publication Number Publication Date
CN110929697A CN110929697A (en) 2020-03-27
CN110929697B true CN110929697B (en) 2021-04-13

Family

ID=69863997

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911301003.8A Active CN110929697B (en) 2019-12-17 2019-12-17 Neural network target identification method and system based on residual error structure

Country Status (1)

Country Link
CN (1) CN110929697B (en)

Families Citing this family (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111401297A (en) * 2020-04-03 2020-07-10 天津理工大学 Triphibian robot target recognition system and method based on edge calculation and neural network
CN111652059B (en) * 2020-04-27 2023-03-24 西北大学 Target identification model construction and identification method and device based on computational ghost imaging
CN111596269B (en) * 2020-05-25 2022-04-19 中国人民解放军海军航空大学 Method for detecting radar composite detection target capability
CN111860290B (en) * 2020-07-16 2022-09-20 大连理工大学 Method for identifying cracks in laser cladding process
CN112307984B (en) * 2020-11-02 2023-02-17 安徽工业大学 Safety helmet detection method and device based on neural network
CN112508625B (en) * 2020-12-18 2022-10-21 国网河南省电力公司经济技术研究院 Intelligent inspection modeling method based on multi-branch residual attention network
CN113240081B (en) * 2021-05-06 2022-03-22 西安电子科技大学 High-resolution range profile target robust identification method aiming at radar carrier frequency transformation
CN113159218A (en) * 2021-05-12 2021-07-23 北京联合大学 Radar HRRP multi-target identification method and system based on improved CNN
CN113297986A (en) * 2021-05-27 2021-08-24 新东方教育科技集团有限公司 Handwritten character recognition method, device, medium and electronic equipment
CN113298092B (en) * 2021-05-28 2024-06-28 有米科技股份有限公司 Neural network training method and device for extracting multi-level image contour information
CN113159078B (en) * 2021-06-04 2022-10-04 中国人民解放军陆军工程大学 Image data identification system and method based on neural network
CN113762479A (en) * 2021-09-10 2021-12-07 深圳朴生智能科技有限公司 Neural network optimization method and device
CN114022745A (en) * 2021-11-05 2022-02-08 光大科技有限公司 Neural network model training method and device
CN114022671B (en) * 2021-11-11 2024-06-21 中国矿业大学 Intelligent cableway cable position identification method based on residual neural network
CN114366047B (en) * 2022-01-27 2023-05-09 上海国民集团健康科技有限公司 Multi-task neural network pulse condition data processing method, system and terminal
CN114973207B (en) * 2022-08-01 2022-10-21 成都航空职业技术学院 Road sign identification method based on target detection
CN115830633B (en) * 2022-11-24 2023-07-25 之江实验室 Pedestrian re-recognition method and system based on multi-task learning residual neural network

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108710826A (en) * 2018-04-13 2018-10-26 燕山大学 A kind of traffic sign deep learning mode identification method
CN108921029A (en) * 2018-06-04 2018-11-30 浙江大学 A kind of SAR automatic target recognition method merging residual error convolutional neural networks and PCA dimensionality reduction
CN109086799A (en) * 2018-07-04 2018-12-25 江苏大学 A kind of crop leaf disease recognition method based on improvement convolutional neural networks model AlexNet
CN109116203A (en) * 2018-10-31 2019-01-01 红相股份有限公司 Power equipment partial discharges fault diagnostic method based on convolutional neural networks
CN109375186A (en) * 2018-11-22 2019-02-22 中国人民解放军海军航空大学 Radar target identification method based on the multiple dimensioned one-dimensional convolutional neural networks of depth residual error
CN109389045A (en) * 2018-09-10 2019-02-26 广州杰赛科技股份有限公司 Micro- expression recognition method and device based on mixing space-time convolution model
CN110472564A (en) * 2019-08-14 2019-11-19 成都中科云集信息技术有限公司 A kind of micro- Expression Recognition depression method of two-way LSTM based on feature pyramid network

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107194371B (en) * 2017-06-14 2020-06-09 易视腾科技股份有限公司 User concentration degree identification method and system based on hierarchical convolutional neural network
US11049018B2 (en) * 2017-06-23 2021-06-29 Nvidia Corporation Transforming convolutional neural networks for visual sequence learning
US10423861B2 (en) * 2017-10-16 2019-09-24 Illumina, Inc. Deep learning-based techniques for training deep convolutional neural networks
CN109285112A (en) * 2018-09-25 2019-01-29 京东方科技集团股份有限公司 Image processing method neural network based, image processing apparatus
CN109472240B (en) * 2018-11-12 2020-02-28 北京影谱科技股份有限公司 Face recognition multi-model adaptive feature fusion enhancement method and device

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108710826A (en) * 2018-04-13 2018-10-26 燕山大学 A kind of traffic sign deep learning mode identification method
CN108921029A (en) * 2018-06-04 2018-11-30 浙江大学 A kind of SAR automatic target recognition method merging residual error convolutional neural networks and PCA dimensionality reduction
CN109086799A (en) * 2018-07-04 2018-12-25 江苏大学 A kind of crop leaf disease recognition method based on improvement convolutional neural networks model AlexNet
CN109389045A (en) * 2018-09-10 2019-02-26 广州杰赛科技股份有限公司 Micro- expression recognition method and device based on mixing space-time convolution model
CN109116203A (en) * 2018-10-31 2019-01-01 红相股份有限公司 Power equipment partial discharges fault diagnostic method based on convolutional neural networks
CN109375186A (en) * 2018-11-22 2019-02-22 中国人民解放军海军航空大学 Radar target identification method based on the multiple dimensioned one-dimensional convolutional neural networks of depth residual error
CN110472564A (en) * 2019-08-14 2019-11-19 成都中科云集信息技术有限公司 A kind of micro- Expression Recognition depression method of two-way LSTM based on feature pyramid network

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Deep Residual Learning for Image Recognition;Kaiming He 等;《2016 IEEE Conference on Computer Vision and Pattern Recognition》;20161231;第770-778页 *
Radar HRRP Target Recognition Based on Deep One-Dimensional Residual-Inception Network;Chen GUO 等;《IEEE Access》;20190110;第9191-9204页 *
基于分组残差结构的轻量级卷积神经网络设计;李鹏 等;《微电子学与计算机》;20190731;第36卷(第7期);第43-47页 *

Also Published As

Publication number Publication date
CN110929697A (en) 2020-03-27

Similar Documents

Publication Publication Date Title
CN110929697B (en) Neural network target identification method and system based on residual error structure
EP3908943B1 (en) Method, apparatus, electronic device and computer readable storage medium for image searching
CN107945204B (en) Pixel-level image matting method based on generation countermeasure network
CN106845529B (en) Image feature identification method based on multi-view convolution neural network
CN111126256B (en) Hyperspectral image classification method based on self-adaptive space-spectrum multi-scale network
CN105701502B (en) Automatic image annotation method based on Monte Carlo data equalization
Qi Hierarchically gated deep networks for semantic segmentation
JP4618098B2 (en) Image processing system
KR20200023266A (en) Online progressive real-time learning to tag and label data streams for deep neural networks and neural network applications
CN110533024B (en) Double-quadratic pooling fine-grained image classification method based on multi-scale ROI (region of interest) features
CN108875076B (en) Rapid trademark image retrieval method based on Attention mechanism and convolutional neural network
CN112561027A (en) Neural network architecture searching method, image processing method, device and storage medium
CN109840518B (en) Visual tracking method combining classification and domain adaptation
Bai et al. Coordinate CNNs and LSTMs to categorize scene images with multi-views and multi-levels of abstraction
CN111160225B (en) Human body analysis method and device based on deep learning
CN112308115A (en) Multi-label image deep learning classification method and equipment
CN110765960B (en) Pedestrian re-identification method for adaptive multi-task deep learning
CN111274422A (en) Model training method, image feature extraction method and device and electronic equipment
CN113159067A (en) Fine-grained image identification method and device based on multi-grained local feature soft association aggregation
CN111311702A (en) Image generation and identification module and method based on BlockGAN
CN110473195A (en) It is a kind of can automatic customization medicine lesion detection framework and method
CN111400525A (en) Intelligent fashionable garment matching and recommending method based on visual combination relation learning
CN115131613A (en) Small sample image classification method based on multidirectional knowledge migration
CN115880027A (en) Electronic commerce website commodity seasonal prediction model creation method
CN115116139A (en) Multi-granularity human body action classification method based on graph convolution network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant