Signal optimization preprocessing-based adolescent schizophrenia deep learning recognition system and method

Abstract

The invention provides a signal optimization preprocessing-based adolescent schizophrenia deep learning recognition system and method. The system comprises a data preprocessing module, a data fusion module and a schizophrenia recognition module based on a deep learning model which are arranged in sequence, the data preprocessing module is used for carrying out filtering processing on the screened electroencephalogram signals, retaining frequency bands which are most obviously represented by schizophrenia, extracting power characteristics of each frequency band through a time-frequency analysis method, and carrying out compression and normalization processing; the data fusion module is used for disrupting and fusing a plurality of electroencephalogram signal data sets; and the schizophrenia identification module based on the deep learning model is used for realizing accurate identification of the schizophrenia patient in combination with the deep learning model. By optimizing the data processing flow and the model training method, the schizophrenia identification accuracy and stability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of schizophrenia recognition, and in particular to a deep learning recognition system and method for adolescent schizophrenia based on signal optimization preprocessing. Background Art

[0002] In recent years, the rapid development of neuroimaging technologies, such as brain magnetic resonance imaging (MRI), functional magnetic resonance imaging (fMRI), and electroencephalography (EEG), has provided researchers with an important means to obtain brain structure and functional image data. These imaging data have played an important role in the identification and diagnosis of mental illness, especially in the study of adolescent schizophrenia. By analyzing MRI images, fMRI data, and other neuroimaging data, researchers are able to explore the association between abnormal brain structure or function and schizophrenia. These studies help identify changes in brain regions associated with schizophrenia, thereby providing a scientific basis for early warning of the disease.

[0003] In the process of medical image analysis, a deep understanding of electroencephalogram (EEG) signals is crucial. FFT, or Fast Fourier Transform, is an algorithm for efficiently calculating discrete Fourier transforms. The FFT spectrogram generated based on this is very important in the field of signal processing. It can convert time domain signals into frequency domain signals and decompose complex time domain signals into superpositions of sine and cosine waves of different frequencies. The meaning of each parameter is clear: the horizontal axis represents the change of the signal in time in seconds, and the length of the time axis depends on the duration of the input data; the vertical axis displays the different frequency components contained in the signal in Hertz; the color intensity (power spectrum density) is in dB / Hz, and 10*log10(Sxx) is used for decibel conversion. The 'inferno' color map is used, and dark colors (blue / purple) have low energy, and bright colors (yellow / white) have high energy. FFT spectrograms are widely used. In the medical field, they can assist in analyzing EEG and ECG signals for disease diagnosis.

[0004] In the field of medical image analysis, convolutional neural network (CNN), as a deep learning model, has attracted widespread attention due to its outstanding performance in image recognition and processing. CNN can automatically extract features from a large amount of image data, avoiding the tedious process of manually designing features in traditional methods, thus showing great potential in medical image analysis. Previous studies have shown that CNN can effectively process brain MRI or fMRI images and identify features related to mental illness (including schizophrenia). For example, by analyzing images of abnormal brain structure or function through CNN, brain lesions in patients with schizophrenia, such as gray matter reduction, white matter lesions, and abnormal connectivity between brain regions, can be identified. In the diagnosis of schizophrenia, CNN can not only improve the accuracy of image data analysis, but also reduce human errors and improve the reliability of diagnosis. In addition, the CNN model can be trained with a large amount of labeled data to gradually learn the brain image features of patients with different types of schizophrenia, thus having strong adaptive capabilities.

[0005] EEG signals are mainly divided into the following bands: Delta wave (δ wave): frequency 0.5-4Hz, usually associated with deep sleep; Theta wave (θ wave): frequency 4-8Hz, schizophrenia patients often show an increase in θ waves, especially in the awake state, the abnormal increase of θ waves may be related to cognitive impairment, attention problems and perceptual abnormalities; Alpha wave (α wave): frequency 8-13Hz, schizophrenia patients usually show a decrease or abnormal change in α waves. Especially when the patient closes his eyes, the disappearance or instability of α waves may be related to attention deficits, emotional regulation problems and cognitive dysfunction; Beta wave (β wave): frequency 13-30Hz, schizophrenia patients may also show an abnormal increase in β waves, especially in the resting state, which may be related to symptoms related to excessive excitability, anxiety or movement; Gamma wave (γ wave): frequency 30-100Hz, studies have found that the γ waves of schizophrenia patients may be weakened or less synchronized, especially when performing cognitive tasks, which is greatly related to information processing and memory disorders.

[0006] However, the pathological mechanism of schizophrenia is complex, and a single neuroimaging data may not fully reflect the characteristics of the disease. Therefore, researchers began to explore methods of combining multimodal data (such as structural MRI, functional MRI, and EEG) with CNN. By comprehensively analyzing multiple types of imaging data, the complexity of the disease can be captured more comprehensively, thereby further improving the accuracy of recognition and the comprehensiveness of diagnosis. This method of multimodal data fusion provides new research directions and technical support for the early diagnosis and precision treatment of schizophrenia. Summary of the invention

[0007] In order to achieve the above objectives, the present invention provides a deep learning identification system and method for adolescent schizophrenia based on signal optimization preprocessing, which can realize accurate identification of schizophrenia patients through optimized data processing flow and model training method, and has strong robustness and generalization ability.

[0008] The specific plan is as follows:

[0009] The deep learning recognition system for adolescent schizophrenia based on signal optimization preprocessing includes a data preprocessing module, a data fusion module and a schizophrenia recognition module based on a deep learning model, which are arranged in sequence; the data preprocessing module is used to filter the screened EEG signals, retain the frequency bands with the most obvious manifestations of schizophrenia, and extract the power characteristics of each frequency band through a time-frequency analysis method, and perform compression and normalization processing; the data fusion module is used to shuffle and fuse multiple EEG signal data sets; the schizophrenia recognition module based on a deep learning model is used to combine the deep learning model to achieve accurate recognition of schizophrenia patients.

[0010] Furthermore, in the data preprocessing module, the screened EEG signals are filtered, specifically including: retaining 16 lead channels including F7, F3, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, and O2, which can effectively reduce the amount of data. These channels are the EEG parts that distinguish schizophrenia patients from ordinary subjects with the highest degree of distinction. By performing a 0.5-30Hz bandpass filtering process on the EEG signal, high-frequency noise and low-frequency interference signals are removed, and only the frequency bands with the most obvious manifestations of schizophrenia are retained. This filtering process can significantly improve the quality of EEG signals, thereby improving the accuracy of identifying schizophrenia patients. The key point of this technology is to improve the signal-to-noise ratio, enhance the effectiveness of the signal, reduce the training cost, and improve the recognition accuracy through specific frequency band filtering optimization to provide a high-quality data basis for subsequent analysis.

[0011] Furthermore, the time-frequency analysis method is to perform FFT spectrogram transformation on the preprocessed EEG signal to generate spectrograms corresponding to different leads as training input of the deep learning model; wherein, the FFT spectrogram directly extracts the power characteristics of each frequency band, making it easier for the model to capture frequency domain patterns that are strongly correlated with the disease, and compress the data into low-dimensional frequency domain features, thereby reducing computational complexity and the risk of overfitting, and normalizing the frequency band energy to reduce interference from individual physiological differences and improve the model's generalization ability.

[0012] Furthermore, the time-frequency analysis method specifically includes: generating a horizontal axis in seconds to represent the temporal changes of the signal, and the length of the time axis depends on the duration of the input data; the vertical axis is in Hertz to display the different frequency components contained in the signal; the color intensity (power spectrum density) is in dB / Hz, 10*log10(Sxx) is used for decibel conversion, and an 'inferno' color map is used, in which dark colors (blue / purple) have low energy and bright colors (yellow / white) have high energy in an FFT spectrum.

[0013] Furthermore, the data fusion module specifically includes: shuffling and fusing multiple resting-state EEG signal data sets to construct a more representative and diverse training data set. In this way, even if the data sources are different, it is still possible to effectively identify whether the tester is a schizophrenia patient. The key point of this technology is to enhance the diversity and representativeness of the data set through data fusion, enhance the generalization ability of the model, enable it to adapt to data from different sources, and maintain a high recognition accuracy.

[0014] Furthermore, the preprocessing strategy and training strategy of the deep learning model specifically include: in terms of data specification, the image size is uniformly scaled to 224×224 pixels, and Rescaling (1. / 255) maps the pixel value to the [0,1] interval (normalization). In terms of performance optimization, .cache() stores the data set into memory to accelerate reading, .prefetch (buffer_size = AUTOTUNE) realizes the pipeline parallelization of data loading and model calculation, and Batch size = 128 balances memory efficiency and gradient stability.

[0015] Furthermore, the schizophrenia recognition module based on the deep learning model specifically includes: based on data preprocessing and fusion, combined with a deep learning model (such as a convolutional neural network CNN or other machine learning algorithms) to achieve accurate recognition of schizophrenia patients. The key point of this technology is to ensure that the system can stably and accurately identify schizophrenia patients in different data sets and scenarios by optimizing the data processing process and model training method.

[0016] The deep learning recognition method for adolescent schizophrenia based on signal optimization preprocessing includes the following steps:

[0017] S1. Extract data from multiple public schizophrenia EEG datasets to obtain a dataset;

[0018] S2. First, the formats of different data sets are preprocessed in a unified manner to ensure data consistency, and then a comprehensive training and validation data set is generated; during the preprocessing process, 16 specific EEG lead channels are selected: F7, F3, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, and O2. These channels can cover the main areas of the brain and help improve the accuracy of identifying adolescent schizophrenia. The data of each subject is standardized, and 1 minute of resting EEG signals are uniformly intercepted; these signals are then resampled to 128Hz to ensure efficient data processing and transmission, and low-frequency and high-frequency noise are removed through 0.5-30Hz bandpass filtering to retain the most clinically significant signal components; finally, the EEG signals of each channel of each subject are converted into frequency domain information to generate a spectrum diagram, namely an FFT spectrum diagram, so that the data features of each subject are represented in a more meaningful way for input into the deep learning model;

[0019] S3. Use a deep learning model based on convolutional neural network to classify and identify EEG signals. 80% of the data set is used for training and 20% for validation. The model is trained using cross-validation to ensure that the model has good generalization ability.

[0020] S4. Use the trained deep learning model to identify and classify adolescent schizophrenia patients and evaluate its performance indicators, including recognition accuracy, loss rate, etc. Draw the receiver operating characteristic ROC curve to further evaluate the sensitivity and specificity of the model to ensure the efficiency and accuracy of the model in early identification of schizophrenia patients.

[0021] Furthermore, the deep learning models in step S3 include: CNN model, VGG-16 model and ChronoNet model. The CNN model uses repeated convolution blocks to enhance local feature extraction, progressive feature map channel expansion (32→128), and double fully connected layers to achieve high-order nonlinear combination. The VGG-16 model uses standardized 3×3 convolution kernel stacking, 5 times spatial downsampling (final feature Figure 7×7), giant fully connected layer (4096 nodes). ChronoNet model uses pyramid channel expansion (32→256), four-stage spatial downsampling (MaxPooling×4), and fully connected layers as feature aggregators. The optimizer is Adam. CNN model: This model can efficiently extract spatial and temporal features from EEG spectrograms, which helps to capture subtle changes in the EEG of patients with schizophrenia. VGG-16 model: The classic convolutional neural network structure has been improved to be suitable for efficient feature extraction of EEG data, further improving recognition accuracy. ChronoNet model: This model is specially designed for the characteristics of time series data, and can process the temporal dynamic features in EEG signals, which is expected to improve the model's ability to distinguish schizophrenia.

[0022] The beneficial effects of the present invention are:

[0023] 1. Improve the accuracy and stability of schizophrenia identification by optimizing data processing procedures and model training methods

[0024] 2. Strong data interpretability: FFT spectrogram is used as data, and the model characteristics are consistent with the frequency domain indicators in medical literature. Doctors do not need to understand complex algorithms and can judge whether the model logic is reasonable based on professional knowledge.

[0025] 3. Strong generalization performance: The model has strong adaptability to new data and the model fitting effect is good.

[0026] 4. Strong robustness: The model can perform efficiently when encountering different data sets and handle the interference and noise in the data sets.

[0027] 5. It is both robust and generalizable, and can accurately identify the EEG signals of the tester in a variety of situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a system diagram of the present invention.

[0029] Figure 2 This is an example of generating a spectrogram.

[0030] Figure 3 This is the structural diagram of the CNN model.

[0031] Figure 4 This is the structural diagram of the VGG-16 model.

[0032] Figure 5 This is the structural diagram of the ChronoNet model.

[0033] Figure 6 is the training curve of the CNN model.

[0034] Figure 7 This is the training curve of the VGG-16 model.

[0035] Figure 8 is the training curve of the ChronoNet model.

[0036] Fig. 9 Evaluate the data for each model.

[0037] Fig.10 This is the ROC evaluation curve of the CNN model.

[0038] Fig.11 This is the ROC evaluation curve of the VGG-16 model.

[0039] Fig.12 ROC evaluation curve for the ChronoNet model DETAILED DESCRIPTION

[0040] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0041] As shown in the figure, the present invention provides a deep learning identification system for adolescent schizophrenia based on signal optimization preprocessing, including a data preprocessing module, a data fusion module and a schizophrenia identification module based on a deep learning model, which are arranged in sequence; the data preprocessing module is used to filter the screened EEG signals, retain the frequency bands with the most obvious manifestations of schizophrenia, and extract the power characteristics of each frequency band through a time-frequency analysis method, and perform compression and normalization processing; the data fusion module is used to shuffle and fuse multiple EEG signal data sets; the schizophrenia identification module based on a deep learning model is used to combine the deep learning model to achieve accurate identification of schizophrenia patients.

[0042] In this embodiment, in the data preprocessing module, the screened EEG signals are filtered, specifically including: retaining 16 lead channels including F7, F3, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, and O2, which can effectively reduce the amount of data. These channels are the EEG parts with the highest distinction between schizophrenia patients and ordinary subjects. By performing a 0.5-30Hz bandpass filtering process on the EEG signal, high-frequency noise and low-frequency interference signals are removed, and only the frequency band with the most obvious manifestation of schizophrenia is retained. This filtering process can significantly improve the quality of EEG signals, thereby improving the accuracy of identifying schizophrenia patients. The key point of this technology is to improve the signal-to-noise ratio, enhance the effectiveness of the signal, reduce the training cost, and improve the recognition accuracy through specific frequency band filtering optimization to provide a high-quality data basis for subsequent analysis.

[0043] In this embodiment, the time-frequency analysis method is to perform FFT spectrogram transformation on the preprocessed EEG signal to generate spectrograms corresponding to different leads as training input of the deep learning model; wherein, the FFT spectrogram directly extracts the power characteristics of each frequency band, making it easier for the model to capture frequency domain patterns that are strongly correlated with the disease, and compress the data into low-dimensional frequency domain features, thereby reducing computational complexity, reducing the risk of overfitting, and normalizing the frequency band energy, reducing interference from individual physiological differences, and improving the model's generalization ability.

[0044] In this embodiment, the time-frequency analysis method specifically includes: generating a horizontal axis in seconds to represent the temporal changes of the signal, and the length of the time axis depends on the duration of the input data; the vertical axis is in Hertz to display the different frequency components contained in the signal; the color intensity (power spectrum density) is in dB / Hz, 10*log10(Sxx) is used for decibel conversion, and an 'inferno' color map is used, in which dark colors (blue / purple) have low energy and bright colors (yellow / white) have high energy in an FFT spectrum.

[0045] In this embodiment, the data fusion module specifically includes: shuffling and fusing multiple resting-state EEG signal data sets to construct a more representative and diverse training data set. In this way, even if the data sources are different, it is still possible to effectively identify whether the tester is a schizophrenia patient. The key point of this technology is to enhance the diversity and representativeness of the data set through data fusion, enhance the generalization ability of the model, enable it to adapt to data from different sources, and maintain a high recognition accuracy.

[0046] In this embodiment, the schizophrenia recognition module based on the deep learning model specifically includes: based on data preprocessing and fusion, combining a deep learning model (such as a convolutional neural network CNN or other machine learning algorithms) to achieve accurate recognition of schizophrenia patients. The key point of this technology is to ensure that the system can stably and accurately identify schizophrenia patients in different data sets and scenarios by optimizing the data processing process and model training method.

[0047] The deep learning recognition method for adolescent schizophrenia based on signal optimization preprocessing includes the following steps:

[0048] S1. Extract data from multiple public schizophrenia EEG datasets to obtain a dataset;

[0049] S2. First, the formats of different data sets are preprocessed in a unified manner to ensure data consistency, and then a comprehensive training and validation data set is generated; during the preprocessing process, 16 specific EEG lead channels are selected: F7, F3, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, and O2. These channels can cover the main areas of the brain and help improve the accuracy of identifying adolescent schizophrenia. The data of each subject is standardized, and 1 minute of resting EEG signals are uniformly intercepted; these signals are then resampled to 128Hz to ensure efficient data processing and transmission, and low-frequency and high-frequency noise are removed through 0.5-30Hz bandpass filtering to retain the most clinically significant signal components; finally, the EEG signals of each channel of each subject are converted into frequency domain information to generate a spectrum diagram, namely an FFT spectrum diagram, so that the data features of each subject are represented in a more meaningful way for input into the deep learning model;

[0050] S3. Use a deep learning model based on convolutional neural network to classify and identify EEG signals. 80% of the data set is used for training and 20% for validation. The model is trained using cross-validation to ensure that the model has good generalization ability.

[0051] S4. Use the trained deep learning model to identify and classify adolescent schizophrenia patients and evaluate its performance indicators, including recognition accuracy, loss rate, etc. Draw the receiver operating characteristic ROC curve to further evaluate the sensitivity and specificity of the model to ensure the efficiency and accuracy of the model in early identification of schizophrenia patients.

[0052] In this embodiment, the deep learning model in step S3 includes: a CNN model, a VGG-16 model and a ChronoNet model. CNN model: This model can efficiently extract spatial and temporal features from the EEG spectrogram, which helps to capture subtle changes in the EEG of schizophrenia patients. VGG-16 model: The classic convolutional neural network structure has been improved to be suitable for efficient feature extraction of EEG data, further improving recognition accuracy. ChronoNet model: This model is specially designed for the characteristics of time series data, and can process the temporal dynamic features in EEG signals, which is expected to improve the model's ability to distinguish schizophrenia.

[0053] In the present invention, the preprocessing enables the trained CNN model to achieve an accuracy of 99.12% and a loss value of 0.0247 in the evaluation, and the fitting is very good; the ChronoNet model can achieve an accuracy of 99.12% and a loss value of 0.0302 in the evaluation, and the effect is also good; the VGG-16 model can achieve an accuracy of 98.68% and a loss value of 0.0601 in the evaluation. The model with a simple structure achieves a higher accuracy and a lower loss value for identifying schizophrenia patients under this preprocessing, saving training costs.

[0054] The above is a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not deviate from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A deep learning recognition system for adolescent schizophrenia based on signal optimization preprocessing, characterized by: It includes a data preprocessing module, a data fusion module and a schizophrenia recognition module based on a deep learning model, which are arranged in sequence; the data preprocessing module is used to filter the screened EEG signals, retain the frequency bands with the most obvious manifestations of schizophrenia, and extract the power characteristics of each frequency band through a time-frequency analysis method, and perform compression and normalization processing; the data fusion module is used to shuffle and fuse multiple EEG signal data sets; the schizophrenia recognition module based on a deep learning model is used to combine the deep learning model to achieve accurate identification of schizophrenia patients.

2. The deep learning recognition system for adolescent schizophrenia based on signal optimization preprocessing according to claim 1 is characterized in that: In the data preprocessing module, the screened EEG signals are filtered, specifically including: retaining a total of 16 lead channels, including F7, F3, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, and O2, and performing a 0.5-30 Hz bandpass filtering process on the EEG signals to remove high-frequency noise and low-frequency interference signals, and only retaining the frequency band where schizophrenia is most obviously manifested.

3. The deep learning recognition system for adolescent schizophrenia based on signal optimization preprocessing according to claim 1 is characterized in that: The time-frequency analysis method is to transform the preprocessed EEG signal into an FFT spectrogram to generate spectrograms corresponding to different leads as training inputs for the deep learning model; wherein the FFT spectrogram directly extracts the power features of each frequency band, compresses the data into low-dimensional frequency domain features, and normalizes the frequency band energy.

4. The deep learning recognition system for adolescent schizophrenia based on signal optimization preprocessing according to claim 3 is characterized in that: The time-frequency analysis method specifically includes: generating a horizontal axis in seconds to represent the temporal changes of the signal, and the length of the time axis depends on the duration of the input data; the vertical axis in Hertz to display the different frequency components contained in the signal; the color intensity in dB / Hz, using 10*log10 for decibel conversion, using an 'inferno' color map, and an FFT spectrum map with low energy in dark colors and high energy in bright colors.

5. The deep learning recognition system for adolescent schizophrenia based on signal optimization preprocessing according to claim 1 is characterized in that: The data fusion module specifically includes: shuffling and fusing multiple resting-state EEG signal data sets to construct a more representative and diverse training data set.

6. The deep learning recognition system for adolescent schizophrenia based on signal optimization preprocessing according to claim 1 is characterized in that: The preprocessing strategy and training strategy of the deep learning model: in terms of data specification, the image size is uniformly scaled to 224×224 pixels, and Rescaling (1. / 255) maps the pixel value to the [0,1] interval; in terms of performance optimization, .cache() stores the data set into memory to accelerate reading, .prefetch (buffer_size = AUTOTUNE) realizes the pipeline parallelism of data loading and model calculation, and Batch size = 128 balances memory efficiency and gradient stability.

7. The deep learning recognition system for adolescent schizophrenia based on signal optimization preprocessing according to claim 1 is characterized in that: The schizophrenia recognition module based on the deep learning model specifically includes: based on data preprocessing and fusion, combining the deep learning model to achieve accurate recognition of schizophrenia patients.

8. A deep learning recognition method for adolescent schizophrenia based on signal optimization preprocessing, using a system as described in any one of claims 1 to 7, characterized in that: The following steps are involved: S1. Extract data from multiple public schizophrenia EEG datasets to obtain a dataset; S2. First, the formats of different data sets are preprocessed in a unified manner to ensure data consistency, and then a comprehensive training and validation data set is generated. During the preprocessing process, 16 specific EEG lead channels are selected: F7, F3, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, O2, and the data of each subject is standardized, and 1 minute of resting EEG signals are uniformly intercepted. These signals are then resampled to 128Hz to ensure efficient data processing and transmission, and low-frequency and high-frequency noise are removed through 0.5-30Hz bandpass filtering to retain the most clinically significant signal components. Finally, the EEG signals of each channel of each subject are converted into frequency domain information to generate a spectrum diagram, namely, an FFT spectrum diagram. S3. Use a deep learning model based on convolutional neural network to classify and identify EEG signals. 80% of the data set is used for training and 20% for validation. The model is trained using cross-validation to ensure that the model has good generalization ability. S4. Use the trained deep learning model to identify and classify adolescent schizophrenia patients, evaluate its performance indicators, and draw the receiver operating characteristic (ROC) curve to further evaluate the sensitivity and specificity of the model to ensure the efficiency and accuracy of the model in the early identification of schizophrenia patients.

9. The deep learning identification method for adolescent schizophrenia based on signal optimization preprocessing according to claim 8 is characterized in that: The deep learning models in step S3 include: CNN model, VGG-16 model and ChronoNet model; wherein the CNN model uses repeated convolution blocks to enhance local feature extraction, progressive feature map channel expansion, 32→128, and double fully connected layers to achieve high-order nonlinear combination; the VGG-16 model uses standardized 3×3 convolution kernel stacking, 5 times spatial downsampling, final feature map 7×7, giant fully connected layer, 4096 nodes; the ChronoNet model uses pyramid channel expansion, 32→256, four-stage spatial downsampling, MaxPooling×4, and a fully connected layer as a feature aggregator; the optimizer is Adam.

Images