A neuron connection method based on discrete sequence floating-point conversion

Abstract

This paper discloses a novel neuron connection method based on floating-point conversion of discrete sequences. The method involves extracting image features through convolution to obtain a feature sequence derived from a large number of convolution kernels. Position-marking and partitioning the feature sequence are performed to construct a 10th-order geometric array and a learnable integer sequence. Using an aggregation approach, the discrete sequence is converted to floating point, reducing the feature sequence's data dimension to one. This method effectively reduces the number of parameters in the fully connected layer, meeting the requirements for lightweight models.

Description

Technical Field

[0001] The present invention relates to the technical field of convolutional neural networks, and in particular to a neuron connection method based on discrete sequence floating-point conversion. Background Art

[0002] Humanity is entering the information age, and artificial intelligence is on the rise. Computer vision, the simulation of biological vision using computers and related equipment, is a key area of ​​artificial intelligence. With the development of computer vision, the use of convolutional neural networks (CNNs) for training models has become a hot topic in the industry. CNNs are the cornerstone of computer vision and can be used in a variety of computer vision tasks, including image classification, object detection, and semantic segmentation. Convolutional neural networks primarily consist of a data input layer, convolutional layers, pooling layers, and fully connected layers. Their network structure typically consists of multiple convolutional and pooling layers, followed by one or more fully connected layers.

[0003] Generally speaking, a fully connected layer is typically found at the end of a convolutional neural network. This layer maps the feature space computed by the previous layers (convolutional and pooling layers) to the sample label space. Each neuron in a fully connected layer is fully connected to all neurons in the previous layer, integrating local information that discriminates between classes from the convolutional or pooling layers and reducing the influence of feature position on the classification results. The activation function for each neuron in a fully connected layer typically uses the Reluctant Unit (ReLU) function, and the output of the final fully connected layer can be classified using SoftMax logistic regression. Due to their fully connected nature, fully connected layers typically have the most parameters. Parameter issues with fully connected layers primarily manifest in two aspects. First, the number of parameters in fully connected layers is excessive, accounting for approximately 80% of the total network parameters, resulting in high computational complexity during training and testing. Second, their redundant parameters can easily lead to overfitting.

[0004] In the agricultural sector, the use of mobile devices for plant disease identification is becoming increasingly important, requiring lightweight network models for mobile devices. However, in existing network models, fully connected layers are unsuitable for use in lightweight models due to their redundant parameters and heavy computational burden. A current solution involves parameter pruning, which involves designing evaluation criteria for network parameters and removing redundant parameters based on these criteria to reduce the computational burden of the fully connected layer. Parameter pruning can reduce the number of neuronal connections and prevent model overfitting, but it can lead to irreparable incompleteness in the network model structure, reducing the performance of plant disease identification models. Therefore, while ensuring the performance of plant disease identification models, reducing the computational burden of the fully connected layer remains a key technical challenge in this field. To address this, CN114049517A from Shihezi University discloses an FPGA-based method for image classification and detection of cotton plants and weeds. This method uses a quantization-aware training method to quantize the network model. A processing element with a large number of multipliers is used for fully connected computations. Input neurons and weights are flattened and concatenated into a customized form, allowing input neurons to be reused during the computation process, reducing model size. Although this method reduces the number of neuronal connections, the overall reduction is not significant, and the number of model parameters is still very large, which cannot meet the requirements of lightweight model. Summary of the Invention

[0005] The purpose of the present invention is to overcome the above shortcomings and provide a new neuron connection method based on discrete sequence floating-point conversion that can effectively reduce the parameters of the fully connected layer and adapt to the lightweight requirements of the model.

[0006] To achieve the above purpose, the technical solution of the present invention is as follows:

[0007] The present invention provides a novel neuron connection method based on discrete sequence floating-point conversion, comprising the following steps:

[0008] (1) obtaining original image data from an image dataset, selecting an existing convolutional neural network model, removing the fully connected layer and prediction layer of the model to obtain a final network model, directly inputting the original image data into the final network model for calculation, and extracting image features through convolution through the convolution layer, activation layer, and pooling layer in the network to obtain a feature sequence extracted by a large number of convolution kernels;

[0009] (2) Partition the feature sequence of length Length, the length of each interval is 4, number each interval according to the starting position, and number the floating-point elements in each interval according to the starting position of the floating-point elements in each interval, numbering them as 0, 1, 2, 3. According to the position number of each floating-point element in the interval, find the position number corresponding to the floating-point element with the largest value in each interval, and reassemble the maximum value number of each interval into an array of length Length / 4, the size is: [1×Length / 4]. Determine whether the length of the array is odd. If the length of the array is odd, skip it. If the length of the array is even, add a node at the end of the array, and the value in the node is 0. So far, the position feature marker sequence α is obtained. n , this step is set as function A(X n ), this step is mathematically expressed as: A(X n )=α n , where: α n is the position feature tag sequence, X n Represents the feature sequence obtained by the original image data through the final network model;

[0010] (3) Construct a 10th-order geometric array β with an odd length n , the length of the 10th-order matrix is ​​the same as α n The length of the array is the same, the value of the middle element of the array is 1, and the elements in the array are represented as β0, β1, β2… …β n , the mathematical expression of this operation is The position feature tag sequence α obtained in step (2) n Perform inner product with the 10th-order geometric matrix to reduce the dimension to 1 and obtain high-dimensional feature floating-point information δ;

[0011] (4) All elements in the learnable integer sequence are integers, ranging from 0 to 9, and are randomly selected. The length of this sequence is the same as the length of the feature mark odd sequence in step (2). This sequence is represented by ε n , change ε n Perform inner product with the 10th-order geometric matrix obtained in step (3) to reduce the dimension to 1 and obtain high-dimensional learnable floating-point information γ;

[0012] (5) Multiply the high-dimensional learnable floating-point information with the high-dimensional feature floating-point information to obtain a dimension of one and a magnitude of 10 x The ultra-high-dimensional floating-point feature information is represented by the symbol Y. This information is aggregated from the feature sequence and has ultra-high-level feature information of the original image data and rich abstract language information.

[0013] (6) Decode the ultra-high-dimensional floating-point feature information Y obtained in step (5) to obtain a decoding sequence μ of length N. The elements in the decoding sequence are represented as μ0, μ1…μ n-1 ,μ n , this step operation is set as function B(Y), and the mathematical expression of this step is B(Y)=μ

[0014]

[0015] Of which: 10 x is the magnitude of the ultra-high-dimensional floating-point feature information Y;

[0016] The above-mentioned novel neuron connection method based on discrete sequence floating-point conversion, wherein: the existing model in step (1) is a classification network model.

[0017] The novel neuron connection method based on discrete sequence floating-point conversion is described above, wherein: the image dataset in step (1) is an existing dataset or based on an image dataset to be extracted. The existing dataset is the ImageNet image dataset.

[0018] The above-mentioned novel neuron connection method based on discrete sequence floating-point conversion, wherein: the integers in the learnable integer sequence described in step (4) are integer-type learnable parameters, which can be updated through network iteration.

[0019] Compared with the existing technology, the present invention has significant beneficial effects. From the above technical solutions, it can be seen that the present invention optimizes the fully connected neuron method with too many parameters in the convolutional neural network. By extracting the features of the image through convolution, a feature sequence extracted by a large number of convolution kernels is obtained. The feature sequence is position-marked and partitioned, and a 10th-order geometric array and a learnable integer sequence are constructed. Using the idea of ​​aggregation, the discrete sequence is converted to floating point, and the data dimension of the feature sequence is reduced to one, thereby effectively reducing the computing power requirements of the existing neural network model and improving the device migration capability of the neural network. . It can effectively reduce the parameters of the fully connected layer and adapt to the lightweight model. The present invention can be used in mobile devices for identifying plant diseases, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the present invention;

[0021] Figure 2 Schematic diagram of the 10th-order matrix of the present invention. DETAILED DESCRIPTION

[0022] The following, in conjunction with the accompanying drawings and preferred embodiments, details on the specific implementation, structure, features and efficacy of a neuron connection method based on discrete sequence floating-point conversion proposed in the present invention are described below.

[0023] Example 1

[0024] See also Figure 1 The present invention provides a neuron connection method based on discrete sequence floating point conversion, comprising the following steps:

[0025] (1) Obtaining original image data from the ImageNet image dataset, selecting an existing VGG convolutional neural network model, removing the fully connected layer and prediction layer of the model to obtain a final network model, directly inputting the original image data into the final network model for calculation, and extracting image features through convolution layers, activation layers, and pooling layers in the network to obtain a feature sequence extracted by a large number of convolution kernels;

[0026] (2) Partition the feature sequence of length 4096, the length of each interval is 4, number each interval according to the starting position, and number the floating-point elements in each interval according to the starting position of the floating-point elements in each interval, numbering them as 0, 1, 2, 3. According to the position number of each floating-point element in the interval, find the position number corresponding to the floating-point element with the largest value in each interval, and reassemble the maximum value number of each interval into an array of length 1024, size: [1×1024]. Determine whether the length of the array is odd. If the length of the array is odd, skip it. If the length of the array is even, add a node at the end of the array, and the value in the node is 0. So far, the position feature marker sequence α is obtained. n , this step is set as function A(X n ), this step is mathematically expressed as: A(X n )=α n , where: α n is the position feature tag sequence, X n Represents the feature sequence obtained by the original image data through the final network model;

[0027] (3) See Figure 2 , construct a 10-order geometric array β with an odd length n , the length of the 10th-order matrix is ​​the same as α n The length of the array is the same, the value of the middle element of the array is 1, and the elements in the array are represented as β0, β1, β2… …β n , the mathematical expression of this operation is The position feature tag sequence α obtained in step (2) nPerform inner product with the 10th-order geometric matrix to reduce the dimension to 1 and obtain high-dimensional feature floating-point information δ;

[0028] (4) All elements in the learnable integer sequence are integers (integer-type learnable parameters that can be updated through network iterations), and the integer value range is 0 to 9, and the value is randomly selected. The length of this sequence is the same as the length of the feature mark odd sequence in step (3), and this sequence is represented by ε n , and ε n Perform inner product with the 10th-order geometric matrix obtained in step (2) to reduce the dimension to 1 and obtain high-dimensional learnable floating-point information γ;

[0029] (5) Multiply the high-dimensional learnable floating-point information with the high-dimensional feature floating-point information to obtain a dimension of one and a magnitude of 10 x The ultra-high-dimensional floating-point feature information is represented by the symbol Y. This information is aggregated from the feature sequence and has ultra-high-level feature information of the original image data and rich abstract language information.

[0030] (6) Decode the ultra-high-dimensional floating-point feature information Y obtained in step (5) to obtain a decoding sequence μ with a length of 1000. The elements in the decoding sequence are represented as μ0, μ1…μ n-1 ,μ n , this step operation is set as function B(Y), and the mathematical expression of this step is B(Y)=μ

[0031]

[0032] Of which: 10 x is the magnitude of the ultra-high-dimensional floating-point feature information Y.

[0033] Example 2

[0034] See also Figure 1 The present invention provides a neuron connection method based on discrete sequence floating point conversion, comprising the following steps:

[0035] (1) Obtaining original image data from an apple pest and disease image dataset, selecting an existing Resnet convolutional neural network model, removing the fully connected layer and prediction layer of the model to obtain a final network model, directly inputting the original image data into the final network model for calculation, and extracting image features through convolution through the convolution layer, activation layer, and pooling layer in the network to obtain a feature sequence extracted by a large number of convolution kernels;

[0036] (2) Partition the feature sequence of length 2048, the length of each interval is 4, number each interval according to the starting position, and number the floating-point elements in each interval according to the starting position of the floating-point elements in each interval, numbering them as 0, 1, 2, 3. According to the position number of each floating-point element in the interval, find the position number corresponding to the floating-point element with the largest value in each interval, and reassemble the maximum value number of each interval into an array of length 512, size: [1×512]. Determine whether the length of the array is odd. If the length of the array is odd, skip it. If the length of the array is even, add a node at the end of the array, and the value in the node is 0. So far, the position feature marker sequence α is obtained. n , this step is set as function A(X n ), this step is mathematically expressed as: A(X n )=α n , where: α n is the position feature tag sequence, X n Represents the feature sequence obtained by the original image data through the final network model;

[0037] (3) See Figure 2 , construct a 10-order geometric array β with an odd length n , the length of the 10th-order matrix is ​​the same as α n The length of the array is the same, the value of the middle element of the array is 1, and the elements in the array are represented as β0, β1, β2… …β n , the mathematical expression of this operation is The position feature tag sequence α obtained in step (2) n Perform inner product with the 10th-order geometric matrix to reduce the dimension to 1 and obtain high-dimensional feature floating-point information δ;

[0038] (4) All elements in the learnable integer sequence are integers (integer learnable parameters that can be updated through network iterations). The integer value range is 0 to 9 and is randomly selected. The length of this sequence is the same as the length of the feature mark odd sequence in step (3). This sequence is represented by ε n , and ε n Perform inner product with the 10th-order geometric matrix obtained in step (2) to reduce the dimension to 1 and obtain high-dimensional learnable floating-point information γ;

[0039] (5) Multiply the high-dimensional learnable floating-point information with the high-dimensional feature floating-point information to obtain a dimension of one and a magnitude of 10 x The ultra-high-dimensional floating-point feature information is represented by the symbol Y. This information is aggregated from the feature sequence and has ultra-high-level feature information of the original image data and rich abstract language information.

[0040] (6) Decode the ultra-high-dimensional floating-point feature information Y obtained in step (5) to obtain a decoding sequence μ with a length of 500. The elements in the decoding sequence are represented as μ0, μ1…μ n-1 ,μ n , this step operation is set as function B(Y), and the mathematical expression of this step is B(Y)=μ

[0041]

[0042] Of which: 10 x is the magnitude of the ultra-high-dimensional floating-point feature information Y.

[0043] Example 3

[0044] See also Figure 1 The present invention provides a neuron connection method based on discrete sequence floating point conversion, comprising the following steps:

[0045] (1) Obtaining original image data from a rice fungal disease image dataset, selecting an existing VGG16 convolutional neural network model, removing the fully connected layer and prediction layer of the model to obtain a final network model, directly inputting the original image data into the final network model for calculation, and extracting image features through convolution through the convolution layer, activation layer, and pooling layer in the network to obtain a feature sequence extracted by a large number of convolution kernels;

[0046] (2) Partition the feature sequence of length 4096, the length of each interval is 4, number each interval according to the starting position, and number the floating-point elements in each interval according to the starting position of the floating-point elements in each interval, numbering them as 0, 1, 2, 3. According to the position number of each floating-point element in the interval, find the position number corresponding to the floating-point element with the largest value in each interval, and reassemble the maximum value number of each interval into an array of length 1024, size: [1×1024]. Determine whether the length of the array is odd. If the length of the array is odd, skip it. If the length of the array is even, add a node at the end of the array, and the value in the node is 0. So far, the position feature marker sequence α is obtained. n , this step is set as function A(X n ), this step is mathematically expressed as: A(X n )=α n , where: α n is the position feature tag sequence, X n Represents the feature sequence obtained by the original image data through the final network model;

[0047] (3) See Figure 2 , construct a 10-order geometric array β with an odd length n, the length of the 10th-order matrix is ​​the same as α n The length of the array is the same, the value of the middle element of the array is 1, and the elements in the array are represented as β0, β1, β2… …β n , the mathematical expression of this operation is The position feature tag sequence α obtained in step (2) n Perform inner product with the 10th-order geometric matrix to reduce the dimension to 1 and obtain high-dimensional feature floating-point information δ;

[0048] (4) All elements in the learnable integer sequence are integers (integer learnable parameters that can be updated through network iterations). The integer value range is 0 to 9 and is randomly selected. The length of this sequence is the same as the length of the feature mark odd sequence in step (3). This sequence is represented by ε n , change ε n Perform inner product with the 10th-order geometric matrix obtained in step (2) to reduce the dimension to 1 and obtain high-dimensional learnable floating-point information γ;

[0049] (5) Multiply the high-dimensional learnable floating-point information with the high-dimensional feature floating-point information to obtain a dimension of one and a magnitude of 10 x The ultra-high-dimensional floating-point feature information is represented by the symbol Y. This information is aggregated from the feature sequence and has ultra-high-level feature information of the original image data and rich abstract language information.

[0050] (6) Decode the ultra-high-dimensional floating-point feature information Y obtained in step (5) to obtain a decoding sequence μ with a length of 1000. The elements in the decoding sequence are represented as μ0, μ1…μ n-1 ,μ n , this step operation is set as function B(Y), and the mathematical expression of this step is B(Y)=μ

[0051]

[0052] Of which: 10 x is the magnitude of the ultra-high-dimensional floating-point feature information Y.

[0053] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A novel neuron connection method based on discrete sequence floating-point conversion, comprising the following steps: (1) obtaining original image data from an image dataset, selecting an existing convolutional neural network model, removing the fully connected layer and prediction layer of the model to obtain a final network model, directly inputting the original image data into the final network model for calculation, and extracting image features through convolution through the convolution layer, activation layer, and pooling layer in the network to obtain a feature sequence extracted by a large number of convolution kernels; (2) Partition the feature sequence of length Length, the length of each interval is 4, number each interval according to the starting position, and number the floating-point elements in each interval according to the starting position of the floating-point elements in each interval, numbering them as 0, 1, 2, 3. According to the position number of each floating-point element in the interval, find the position number corresponding to the floating-point element with the largest value in each interval, and reassemble the maximum value number of each interval into an array of length Length / 4, the size is: [1×Length / 4]. Determine whether the length of the array is odd. If the length of the array is odd, skip it. If the length of the array is even, add a node at the end of the array, and the value in the node is 0. So far, the position feature marker sequence α is obtained. n , this step is set as function A(X n ), this step is mathematically expressed as: A(X n )=α n , where: α n is the position feature tag sequence, X n Represents the feature sequence obtained by the original image data through the final network model; (3) Construct a 10th-order geometric array β with an odd length n , the length of the 10th-order matrix is ​​the same as α n The length of the array is the same, the value of the middle element of the array is 1, and the elements in the array are represented as The mathematical expression of this operation is The position feature tag sequence α obtained in step (2) n Perform inner product with the 10th-order geometric matrix to reduce the dimension to 1 and obtain high-dimensional feature floating-point information δ; (4) All elements in the learnable integer sequence are integers, ranging from 0 to 9, and are randomly selected. The length of this sequence is the same as the length of the feature mark odd sequence in step (2). This sequence is represented by ε n , change ε n Perform inner product with the 10th-order geometric matrix obtained in step (3) to reduce the dimension to 1 and obtain high-dimensional learnable floating-point information γ; (5) Multiply the high-dimensional learnable floating-point information with the high-dimensional feature floating-point information to obtain a dimension of one and a magnitude of 10 x The ultra-high-dimensional floating-point feature information is represented by the symbol Y. This information is aggregated from the feature sequence and has ultra-high-level feature information of the original image data and rich abstract language information. (6) Decode the ultra-high-dimensional floating-point feature information Y obtained in step (5) to obtain a decoding sequence μ of length N. The elements in the decoding sequence are represented as μ0, μ1…μ n-1 ,μ n , this step operation is set as function B(Y), and the mathematical expression of this step is B(Y)=μ Of which: 10 x is the magnitude of the ultra-high-dimensional floating-point feature information Y.

2. A novel neuron connection method based on discrete sequence floating-point conversion as claimed in claim 1, wherein: The existing model in step (1) is a classification network model.

3. The novel neuron connection method based on discrete sequence floating-point conversion according to claim 1, wherein: The image dataset in step (1) is an existing dataset or is based on the image dataset to be extracted.

4. A novel neuron connection method based on discrete sequence floating-point conversion as claimed in claim 3, wherein: The existing dataset is the ImageNet image dataset.

5. The novel neuron connection method based on discrete sequence floating-point conversion according to claim 1, wherein: The integers in the learnable integer sequence described in step (4) are integer-type learnable parameters that can be updated through iterations of the network.

Images