Medical image content extraction method based on trainable multilayer RBF-Chaotic neural network

Abstract

The invention discloses a medical image content extraction method based on a trainable multilayer RBF-Chaotic neural network. According to the method, medical image content is extracted based on the trainable multilayer RBF-Chaotic neural network; the trainable multi-layer RBF-Chaotic neural network is obtained by simplifying a LeNet-5 network structure, and comprises the steps of removing all maximum pooling layers in the LeNet-5 network structure, reconstructing a classifier into a three-layer full-connection layer, and introducing chaotic momentum generated by a 2DES function in a training process to replace traditional SGD momentum. According to the method, the model is thinner, the occupation of a video memory is lower, a consumer-level video card can be completely loaded, offline clustering is not needed, the model only changes training and does not change reasoning, hospital night long-time training can be successful at a time, and the manual parameter adjustment cost is remarkably reduced.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of medical image processing and artificial intelligence, and in particular to a method for extracting medical image content based on a trainable multilayer RBF-Chaotic neural network. Background Technology

[0002] In the field of medical image segmentation, deep convolutional neural networks (DCNNs) have become the mainstream method for extracting lesions or organ regions from images such as MRI and CT scans. LeNet-5 is one of the earliest proposed classic DCNN architectures, and its topology is as follows: Figure 1 As shown, a hierarchical design of "two convolutional layers - two max pooling layers - three fully connected layers" is adopted: the convolutional layers extract local features by sliding convolutional kernels; the max pooling layers downsample with a 2×2 window and a stride of 2 to reduce the feature map resolution and the number of parameters; finally, the three fully connected layers output the classification or segmentation results. During the training phase, the backpropagation algorithm is typically used in conjunction with a stochastic gradient descent (SGD) optimizer to minimize the loss function and update the network weights.

[0003] However, when LeNet-5 architectures are applied to medical imaging, the following common problems exist: (1) Limited segmentation accuracy: Medical images have nonlinear features such as uneven gray levels, blurred boundaries and high noise. Traditional linear convolution + ReLU activation function is difficult to fully model, resulting in inaccurate edge segmentation.

[0004] (2) Loss of feature information: Max pooling layers irreversibly discard spatial details while reducing resolution, affecting the network's ability to recognize small structures.

[0005] (3) Low training efficiency: Deep networks have a large number of parameters, high GPU memory usage, and long training time; at the same time, the SGD optimizer is prone to getting trapped in local minima and has a slow convergence speed.

[0006] (4) Insufficient generalization ability: The low-resolution features after pooling are sensitive to image deformation and differences in scanning protocols, and the performance drops significantly during cross-center testing. Summary of the Invention

[0007] To overcome the shortcomings and deficiencies of existing technologies, the purpose of this invention is to provide a method for extracting medical image content based on a trainable multilayer RBF-Chaotic neural network.

[0008] The technical solution provided by this invention is as follows: A method for extracting medical image content, wherein the method is based on a trainable multilayer RBF-Chaotic neural network to extract medical image content; the trainable multilayer RBF-Chaotic neural network is obtained by simplifying the LeNet-5 network structure; the simplified trainable multilayer RBF-Chaotic neural network includes an input layer, a deep RBF neural network, and three fully connected layers; the deep RBF neural network contains three hidden layers; The simplification includes: (1) Remove all max pooling layers from the LeNet-5 network structure; (2) Reconstruct the classifier into a three-layer fully connected layer; (3) During the training process, a two-dimensional exponential-sinusoidal hyperchaotic mapping system 2DES is embedded in the weight update loop, and chaotic momentum is used to replace the traditional SGD momentum.

[0009] Preferably, all neurons in the hidden layers are configured as RBF neurons to replace conventional linear neurons, and the ratio of the number of neurons in the three hidden layers is 1:2:1.

[0010] Preferably, during the training phase, the chaotic perturbation generated by the 2DES is introduced into the multilayer RBF-Chaotic neural network, and specific parameters are set in the 2DES chaotic momentum update unit to dynamically adjust the momentum gradient update direction.

[0011] Preferably, among the specific parameters, the control parameters are a=0.05, b=1.3; the initial values ​​are x0=0.15, y0=0.25.

[0012] Preferably, a structure-level perturbation mechanism is added to the chaotic perturbation generated by the 2DES to protect the training process.

[0013] Preferably, the three-layer fully connected layer structure is 800→500→number of categories.

[0014] Preferably, in the hidden layer, the center parameter, width parameter, output weight, and bias are all designed as trainable parameters and are updated synchronously with subsequent layers through the backpropagation mechanism, forming an end-to-end trainable DeepRBF-5Chaotic network structure, achieving structural-level optimization rather than simple algorithm adjustment.

[0015] Preferably, in the simplification, all spatial features are retained except for the removal of all max-pooling layers in the LeNet-5 network structure, and the number of fully connected synapses is increased.

[0016] Preferably, the simplification further includes: (4) all parameters { c j ,σ j , w j , b End-to-end synchronous updates are achieved through backpropagation.

[0017] This invention overcomes the shortcomings of existing technologies and provides a method for extracting medical image content based on a trainable multilayer RBF-Chaotic neural network, used for pixel-level segmentation of medical images. The technical solution of this invention includes: (1) Feature Extractor Innovation: A trainable feature extraction network composed of a deep RBF (three hidden layer neurons in a ratio of 1:2:1) neural network is adopted, which completely replaces the convolutional and pooling layers in the traditional LeNet-5. (2) No pooling design: The max pooling layer is completely removed from the network to ensure that the feature map resolution is completely preserved from input to feature extraction. (3) Wide channel classifier: The feature vector is first transformed into a high-dimensional nonlinear transformation through a two-layer "wide channel" fully connected module (800 to 500 nodes) to fully retain the feature information and improve the mapping capacity. Finally, the pixel classification is completed through an output layer with the number of nodes equal to the number of categories.

[0018] (4) Embed a two-dimensional exponential-sinusoidal chaotic mapping (2DES) in the weight update loop to replace the traditional SGD momentum; (5) All parameters { c j , σ j , w j , b End-to-end synchronous updates are achieved through backpropagation.

[0019] Parameter settings: The 2DES chaotic momentum update mechanism uses fixed parameters a=0.05, b=1.3. Experiments verify that under these settings, the D5CN network (150 epochs, batch=50) can converge to... The error indicates that the parameter combination is effective and reliable.

[0020] 1. Overall Architecture This invention uses the LeNet-5 network as the basic network architecture, and simplifies the LeNet-5 network structure to take into account the hardware limitations of the running platform.

[0021] Traditional LeNet-5 network architecture, such as Figure 1As shown, it includes an input layer, a first convolutional layer C1, a first pooling layer S2, a third convolutional layer C3, a second pooling layer S4, a first fully connected layer C5, a second fully connected layer F6, and an output layer. Figure 1 As can be seen, the classic LeNet-5 network consists of a stack of "two convolutions, two pooling layers, and three fully connected layers": data is first processed by convolutional kernels to extract local features, then downsampled by max pooling, and the compressed feature map is fed into subsequent convolution-pooling combinations; after two rounds of "extraction-compression", all fine-grained information is flattened and fed into the final three-level fully connected sub-network for final decision-making and outputting the segmentation result. Max pooling reduces the resolution of the feature map, reduces memory usage, and speeds up training, but it also irreversibly discards some spatial details and forces the network to compensate for information loss by increasing the number of fully connected layers, resulting in an increase in overall size and a decrease in efficiency.

[0022] To address the aforementioned shortcomings, this invention innovatively abandons traditional linear convolution and pooling operations, replacing them with a feature extraction engine composed of a deep RBF neural network, constructing a completely new unified kernel architecture. This architecture completely removes the max-pooling layer and reconstructs the classifier into three fully connected layers (800→500→number of classes), improving the model's mapping capacity through a "wide-channel" design. Simultaneously, the training process introduces chaotic momentum generated by the 2DES (two-dimensional exponential-sine) function, replacing the traditional SGD momentum, ultimately forming... Figure 2 The DeepRBF-5Chaotic-Net (D5CN) topology shown includes an input layer → a deep RBF feature extraction network → three fully connected layers (800 → 500 → number of classes).

[0023] In this architecture: The traditional convolution-pooling module has been completely replaced: the original "C1-S2-C3-S4" chain in LeNet-5 has been replaced by three RBF feature extraction layers (through... Figure 4 (The deep RBF neural network is used for feature extraction) has been replaced.

[0024] The fully connected structure has been reconstructed: the three fully connected layers at the end (800→500→number of categories) adopt a "wide channel" design, which is contrary to the LeNet-5 convention of decreasing layer by layer. By broadening the feature pathway, the nonlinear discrimination ability of the model is significantly improved.

[0025] More complete information flow: In the forward path, image features are non-linearly transformed by a deep RBF neural network, with no information downsampling throughout, thus preserving complete details.

[0026] More efficient training: In the reverse path, the 2DES chaotic momentum mechanism is applied to the weight update of the entire network, effectively guiding the optimization process to escape local minima.

[0027] The beneficial effects of this invention after adopting the above technical solution are as follows: (1) Pool-free wide channel topology: zero information loss + drastic reduction in hardware burden This invention completely eliminates max pooling, maintaining the original resolution throughout and preventing premature cropping of edges and details. At the same time, it changes the "deep and narrow" to a "shallow and wide" two-level fully connected layer, concentrating parameters on trainable weights, resulting in a thinner model, lower memory usage, and full loading on consumer-grade graphics cards.

[0028] (2) Multilayer trainable RBF neurons: one-step nonlinear modeling This invention replaces the traditional linear convolution + ReLU with "distance-activation", forming multi-center, multi-scale soft decision boundaries within the same layer, which naturally fits the gray-scale gradation and weak boundaries of medical images; the center, width, and weights are all updated synchronously with BP, eliminating the need for offline clustering and truly achieving end-to-end training.

[0029] (3) 2DES chaotic momentum: structural-level "escape from local minima" mechanism This invention embeds a momentum update loop into a two-dimensional exponential-sinusoidal chaotic mapping, continuously injecting perturbations with random direction and controllable amplitude, allowing the parameters to maintain a non-zero velocity even in flat regions; it only modifies the training and does not modify the inference, enabling successful long-term training in hospitals at night, significantly reducing the cost of manual parameter tuning. Attached Figure Description

[0030] Figure 1 This is a typical LeNet-5 network architecture diagram; Figure 2 This is the overall architecture diagram of the DeepRBF-5Chaotic-Net (D5CN) of this invention; Figure 3 This is the RBF neuron model of the present invention; Figure 4 This is a schematic diagram of the deep RBF feature extraction network of the present invention; Figure 5 This is the training and prediction process of the D5CN network in this invention; Figure 6 This is a flowchart of the D5CN neural network algorithm of this invention; Detailed Implementation

[0031] The technical solution of the present invention will be further described in detail below with reference to specific embodiments, but this does not constitute any limitation on the present invention.

[0032] 1. DeepRBF-5 Chaotic-Net Neural Network (1) RBF radial basis functions RBF (Radial Basis Function) neuron structure: A single neuron can receive signals from multiple signal sources, accumulate these signals through internal accumulation nodes, and then obtain the neuron's output signal by performing a convolution operation on the signals. For example... Figure 3 As shown, the activation function of an RBF neural network consists of a set of radial basis functions. These radial basis functions activate the input vector... With neuron central parameters The distance between (also known as the threshold vector) is used as its independent variable. Its specific form is shown in formula (1): (1) The formula for the radial basis function is shown in formula (2): (2) As can be seen from formula (2), the closer the input variable is to the neuron’s central parameter, the larger the output value of the radial basis function and the maximum output value of the function is 1.

[0033] Traditional Recurrent Block Function (RBF) networks typically employ a three-layer structure of "input-single hidden layer-output," where the number of hidden nodes needs to be manually set in advance, making it difficult to balance approximation accuracy and generalization ability. This invention introduces a three-level hidden layer with a "neuron ratio of 1:2:1" into the RBF framework, forming a unique five-layer RBF topology (input-hidden 1-hidden 2-hidden 3-output). Through a bottleneck mapping of "sparse-expanded-re-sparse," the network achieves a larger effective receptive field and nonlinear characterization ability with the same number of parameters. This deep RBF feature extraction network is structurally divided into five layers. The first layer is the input layer, the second to fourth layers are hidden layers (neuron ratio 1:2:1), and the fifth layer is the output layer. (The text continues with further details about the deep RBF feature extraction network and its structure.) Figure 4 The diagram shown illustrates a deep RBF feature extraction network, including an input layer, hidden layer 1, hidden layer 2, hidden layer 3, and an output layer (out). ij and the weight matrix w ij .

[0034] In the D5CN neural network topology, when the input layer signal is output to the hidden layer, its signal transmission weight is 1. The input signal can be directly transmitted to the hidden layer. When the signal is transmitted through the hidden layer to the output layer, the signal output weight is... .

[0035] The final output signal of the neural network can be obtained from formula (3): (3) The neural network training termination function terminates the training process when the training termination condition is met. Its specific form is shown in formula (4): (4) (2) D5CN network training and prediction process like Figure 5 As shown, the D5CN network consists of training and prediction processes.

[0036] 1) Training phase (closed-loop iteration) ① Input: Registered medical image training samples.

[0037] ② FCM (Fuzzy C-Means) clustering: Perform fuzzy C-means on all pixels to obtain the initial membership matrix U(0), and temporarily assign each pixel i to cluster Kj according to the principle of maximum membership.

[0038] ③ D5CN fitting: Forward: Input the original pixel features (such as gray matter, white matter, cerebrospinal fluid) into D5CN, and output its predicted probabilities for each class. ; Reverse: with The mean squared error of the soft label Y is used as the loss. 2DES chaotic momentum is used to update the weights, and the membership class of pixel i is changed from K... j Change to K n This makes the network output approximate the FCM clustering results.

[0039] ④ Loop termination condition: If the pixel classification change rate is less than ε for two consecutive rounds or the maximum number of iterations is reached, training is stopped and the network parameters are fixed.

[0040] 2) Prediction phase (single feedforward) ① Input: Medical images to be segmented (predicted samples).

[0041] ② Forward computation: Pixels are fed into the trained D5CN network one by one to obtain the final membership probability of each pixel to each class.

[0042] ③ Hard segmentation: Take the category with the highest probability as the label of the pixel and output the segmentation mask of the entire image.

[0043] Through a collaborative mechanism of "FCM clustering to provide coarse distribution → D5CN deep network to refine boundary in a loop", this method makes full use of global gray-level statistical information during the training phase, while relying on three-layer RBF hidden layers and 2DES chaotic momentum update to achieve efficient fitting of local nonlinear boundaries; during the inference phase, only a single feedforward is needed to output high-precision and robust medical image segmentation results, balancing accuracy and real-time performance.

[0044] (3) D5CN network training This invention employs a trainable DeepRBF-5Chaotic-Net (D5CN) structure, where weight parameters are updated via backpropagation and the 2DES chaotic momentum mechanism, such as... Figure 6 As shown, the training process is as follows: 1) Select an appropriate number of T1 structural image data as training samples to form a training sample set. .

[0045] 2) Initialize network parameters: Determine the number of hidden layer nodes During the first round of training, participants were randomly selected. Cluster centers c j , ; Initialize width parameter σ j Output weights w j Bias b = 0, momentum v = 0; Set the maximum number of training rounds.

[0046] 3) Forward propagation For input samples xi ∈R d The output of the input signal after passing through each hidden layer neuron is calculated using the radial basis function (5): (5) Where c j The center of the j-th radial basis function (RBF) is represented.

[0047] The output value of the neural network is calculated using the weighted average according to formula (6).

[0048] (6) Among them, w j b is the output weight, and b is the bias (initial value is 0).

[0049] 4) Calculate the loss (7) Calculate the mean square error between the network output value and the true value. If the training termination condition is met, the training ends; otherwise, continue to the next step of calculation.

[0050] 5) Backpropagation Calculate the loss function L with respect to each training parameter θ gradient ,in θ ∈{ c j , σ j, w j , b}

[0051] 6) 2DES Chaotic Momentum Update Mechanism ① Iteration of chaotic systems The iterative formula of the two-dimensional exponential-sinusoidal hyperchaotic mapping system (2DES) is as follows: (8) Among them, the control parameters (the optimal combination verified by experiments) are: a = 0.05, b = 1.3, initial value: x 0 = 0.15 y 0 = 0.25. The Lyapunov exponent λ > 0 of this system, which exhibits hyperchaotic properties and can generate non-repeating and highly ergodic perturbation sequences, providing diverse directional guidance for momentum updates.

[0052] ② Generation of chaotic perturbation sequences To avoid the instability effects of the initial transient phase of a chaotic system and to ensure the validity and randomness of the perturbation sequence, the generation steps are as follows: Original sequence generation: Run the 2DES system according to the above iterative formula to generate an original chaotic sequence of length N {(x n ,y n )}, n=1,2,…,N; Transient point discarding: The first 15% of transient points are discarded to eliminate the interference of non-chaotic characteristics in the initial stage and retain the subsequent stable segment sequence; Noise enhancement: For the y segment of the stable sequence n Values ​​with Gaussian noise (variance) σ =0.01), further enhancing the randomness and robustness of the perturbation; Perturbation value selection: During each training round, the perturbation value is randomly selected from the processed y using a random index. n Select a value from the sequence as the chaotic perturbation value for the current round (Serial). t (i.e., Serial) t ∈{y n}) ③ Chaotic Momentum Update

[0053] The core of chaotic momentum update is to serialize the chaotic perturbation value. t As part of the dynamic momentum coefficient, the weights of historical momentum are adjusted in real time, thereby introducing controllable randomness into gradient descent and helping the network escape local minima. The update formula is as follows: (9) (10) The parameters are defined as follows: v t : Momentum value in round t (initial v0 = 0); μ: Basic momentum coefficient (fixed at 0.9 to maintain historical momentum inertia); α: Chaotic disturbance intensity coefficient (fixed at 0.01, controlling the amplitude of chaotic influence); Serial t The chaotic perturbation value in round t (taken from the y-value of the 2DES system). n The sequence is typically a small random value with approximately 0.25 ± noise. η: Learning rate (fixed) ); The gradient of the loss function L with respect to the parameter θ θ ∈{ c j , σ j , w j , b}); θ t : Network parameters in round t.

[0054] 7) Return to step 3) and continue until the stopping condition is met.

[0055] After training, the trained neural network is applied to test data to test the effectiveness, generalization and stability of the D5CN neural network algorithm.

[0056] (4) Analysis of D5CN network training results ① Training platform and experimental setup The specific hardware specifications of the D5CN training platform are as follows: CPU: AMD R7-3750H; RAM: 32GB; GPU: NVIDIA 2060 (6GB). Training was performed on the training sample set based on the aforementioned neural network training algorithm. During training, the epoch was set to 150 rounds, the batch size to 50, and the target precision to be 10^-10. -5 The target recognition accuracy was 98%. In comparison, while training D5CN, a regular RBF network was also trained. The training errors of the two neural networks are compared in Tables 1 and 2 below.

[0057] Table 1. Training Error (MSE) of D5CN and Ordinary RBF Networks (5-70 rounds)

[0058] Table 2 Training Error (MSE) of D5CN and Ordinary RBF Network (80~150 rounds)

[0059] The comparison of training errors shown in Tables 1 and 2 reveals that the D5CN network exhibits significantly faster convergence in the early stages of training, with the error approaching zero and remaining stable around the 20th epoch. In contrast, the ordinary RBF network converges slowly, ultimately resulting in a higher error level. These results validate the effectiveness of the proposed 2DES chaotic momentum update mechanism in improving optimization efficiency and avoiding local minima. Furthermore, the Dice coefficients on the validation set remain consistently stable, indicating that the model achieves its desired performance. It does not exhibit overfitting at the order of magnitude accuracy and has good generalization performance.

[0060] The optimized D5CN network can be further deployed on embedded AI computing platforms (such as the NVIDIA Jetson series) to meet the real-time and resource-constrained application requirements of actual medical imaging edge devices through model compression and inference acceleration technologies.

[0061] This invention further provides an implementation process for MRI brain tissue segmentation based on DeepRBF-5Chaotic-Net (D5CN), specifically including: (1) Hardware environment CPU: AMD R7-3750H; RAM: 32 GB; GPU: NVIDIA RTX 2060 (6 GB VRAM).

[0062] Software: Python 3.8 + PyTorch 1.12, FSL 5.0 for image registration.

[0063] (2) Data preparation Open-source MRI database: T1 structural images of 23 volunteers.

[0064] Registration: All images were registered to the MNI152 MNI (Montreal Neurological Institute) standard space using FSL 5.0, with a resampling resolution of 1 mm × 1 mm × 1 mm.

[0065] Division: 20 cases were used for training, and 3 cases were used for testing.

[0066] (3) Network construction Topology: Input layer → Three-layer RBF feature extraction layer → Three-layer fully connected layer (800 → 500 → number of categories).

[0067] Activation: Hidden layers employ Gaussian radial basis functions .

[0068] Chaotic momentum: 2DES mapping parameters a=0.05, b=1.3; initial values =0.15, =0.25; the control parameter remains constant.

[0069] Trainable parameter: center c j Width σ j Output weight w j The bias b is updated synchronously via backpropagation.

[0070] (4) Training process Step 1: Perform fuzzy C-means clustering (FCM) on all pixels in the training set to obtain the initial soft label U(0).

[0071] Step 2: Initialize network parameters and set the learning rate. The basic momentum coefficient μ = 0.9, the chaotic perturbation intensity coefficient α = 0.01, the maximum epoch = 150, and the batch = 50.

[0072] Step 3: Iterative Training a) Forward pass: Pixel grayscale, texture features → D5CN → Output class probability .

[0073] b) Loss: Mean Square Error .

[0074] c) Backwards: Calculate the gradient .

[0075] d) 2DES Chaotic Momentum Update:

[0076] Step 4: Termination condition—pixel classification change rate over two consecutive rounds Or epoch=150, the curing parameter.

[0077] (5) Reasoning process Input the MRI image to be segmented, perform forward calculation pixel by pixel to obtain the membership probability.

[0078] Hard segmentation: Take the category with the highest probability as the final label and output the segmentation mask.

[0079] (6) Data and Indicators Dataset: T1 structural images of 23 volunteers, 20 for training and 3 for testing; Hardware: AMD R7-3750H + 32 GB + RTX 2060 (6 GB); Training configuration: 150 epochs, batch=50, initial learning rate ; Observation metrics: Mean squared error of training set (MSE) and target recognition accuracy.

[0080] (7) Results Training error: D5CN converges to within 150 epochs. However, the error of a regular RBF network in the same round is still higher than that of a standard RBF network. .

[0081] Identification accuracy: D5CN achieves 98%, meeting the clinical requirement of >95%.

[0082] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for extracting medical image content based on a trainable multilayer RBF-Chaotic neural network, characterized in that, The method described above is to extract medical image content based on a trainable multilayer RBF-Chaotic neural network. The trainable multilayer RBF-Chaotic neural network is obtained by simplifying the LeNet-5 network structure. The simplified trainable multilayer RBF-Chaotic neural network includes an input layer, a deep RBF neural network, and three fully connected layers. The deep RBF neural network contains three hidden layers; The simplification includes: (1) Remove all max pooling layers from the LeNet-5 network structure; (2) Reconstruct the classifier into a three-layer fully connected layer; (3) During the training process, a two-dimensional exponential-sinusoidal hyperchaotic mapping system 2DES is embedded in the weight update loop, and chaotic momentum is used to replace the traditional SGD momentum.

2. The method as described in claim 1, characterized in that, Within the hidden layers, all neurons are configured as RBF neurons to replace conventional linear neurons, and the ratio of the number of neurons in the three hidden layers is 1:2:

1.

3. The method as described in claim 2, characterized in that, During the training phase, the chaotic perturbation generated by the 2DES is introduced into the multilayer RBF-Chaotic neural network, and specific parameters are set in the 2DES chaotic momentum update unit to dynamically adjust the momentum gradient update direction.

4. The method as described in claim 3, characterized in that, Among the specific parameters, the control parameters are a=0.05, b=1.3; the initial values ​​are x0=0.15, y0=0.

25.

5. The method as described in claim 3, characterized in that, In the chaotic perturbations generated by the 2DES, an additional structural perturbation mechanism is added to protect the training process.

6. The method as described in claim 1, characterized in that, The three-layer fully connected layer structure is 800→500→number of categories.

7. The method as described in claim 1, characterized in that, In the hidden layer, the center parameter, width parameter, output weight, and bias are all designed as trainable parameters and are updated synchronously with subsequent layers through the backpropagation mechanism, forming an end-to-end trainable DeepRBF-5Chaotic network structure, achieving structural optimization rather than simple algorithm adjustment.

8. The method as described in claim 1, characterized in that, In the simplification, all max-pooling layers in the LeNet-5 network structure are removed, all spatial features are retained, and the number of fully connected synapses is increased.

9. The method as described in claim 1, characterized in that, The simplification also includes: (4) putting all parameters { c j , σ j , w j , b End-to-end synchronous updates are achieved through backpropagation.

Images