A method for underwater acoustic target recognition based on data augmentation and residual CNN

Abstract

This invention relates to an underwater acoustic target recognition method based on data augmentation and residual CNN. It mainly includes extracting MFCC features of underwater acoustic targets, data augmentation processing, and ResNet18 model recognition. It fully leverages the data augmentation advantages of data augmentation technology and the deep feature mining capabilities of the residual CNN model to effectively identify underwater acoustic targets. This study employs the DCGAN model, which has an excellent image generation architecture. The DCGAN discriminator, by incorporating a CNN model, can extract deeper image features, exhibiting significant advantages in image generation and classification. The classification model uses a ResNet18 model with residual connections. This study removes the pooling layer from the ResNet18 model to retain more underwater acoustic feature information and adjusts the network model size to adapt to the feature images input in this study. This research achieves effective underwater acoustic target recognition in scenarios where underwater acoustic target sample data is relatively scarce, utilizing data augmentation techniques and a residual CNN model.

Description

Technical Field

[0001] This invention belongs to the fields of digital signal processing, machine learning and underwater acoustic measurement, and relates to an underwater acoustic target recognition method based on data augmentation and residual CNN. It utilizes data augmentation technology and residual CNN model to realize acoustic signal recognition in scenarios where underwater acoustic target sample data is relatively scarce. Background Technology

[0002] Against the backdrop of the rapid development of machine learning technology, data-driven models such as neural networks can uncover deep features of different target acoustic signals and significantly reduce the impact of noise, effectively achieving autonomous and intelligent classification decisions. Therefore, machine learning methods have been widely researched and applied in the field of sound signal processing. However, the relative scarcity of underwater acoustic target sample data limits the application of machine learning in practical underwater acoustic recognition. In 2013, Kamal et al. applied the Deep Brief Network (DBN) model to the task of passive target recognition of underwater acoustic signals, achieving label-free underwater target acoustic signal recognition (S. Kamal, SK Mohammed, PRSPillai, MH Supriya, “Deep learning architectures for underwater target recognition,”. Sympol 2013, pp. 48-54.). In 2014, Shamir et al., based on the extraction of acoustic features of whales, used machine learning methods to achieve automatic identification of different whale species (L. Shamir, C. Yerby, R. Simpson, A.M. von Benda-Beckmann, P. Tyack, F. Samarra, et al., “Classification of large acoustic datasets using machine learning and crowdsourcing: Application to whale calls,” Journal of the Acoustical Society of America, 2014, vol. 135 No. 2, pp. 953-962.). In 2017, Yue et al. used SVM, DBN, and CNN models to achieve effective identification of ship target acoustic signals (Hao Y, Zhang L, Wang D, et al. The Classification of Underwater Acoustic Targets Based on Deep Learning Methods. 2017 2nd International Conference on Control, Automation and Artificial Intelligence. 2017.).In 2020, Yu et al. built a variety of machine learning models to detect the vocalizations of North Atlantic right whales. The results showed that CNN models could greatly improve accuracy (Y. Shiu, KJPalmer, MARoch, E. Fleishman, X. Liu, EMNosal, T. Helble, D. Cholewiak, D. Gillespie, and H. Klinck, “Deep neural networks for automated detection of marine mammalspecies,” Sci. Rep. 10(1), 1-12 (2020).).

[0003] As the number and complexity of machine learning models increase, the demand for data volume grows ever larger. Only by training models with massive amounts of labeled data can good recognition results be achieved. However, in reality, underwater acoustic target sample data is relatively scarce, limiting the application of machine learning in practical underwater acoustic recognition. To increase the amount of samples required for machine learning training, data augmentation methods have become one of the most popular techniques (Fong, R., & Vedaldi, A. (2019). Occlusions for effective data augmentation in image classification.). Traditional data augmentation techniques expand training samples by adding geometric transformations, color space transformations, etc. However, because traditional data augmentation techniques cannot generate substantial samples and the modifications are relatively small, the training performance after data augmentation is limited. To avoid the limitations of traditional data augmentation techniques, Goodfellow et al. proposed Generative Adversarial Networks (GANs), which generate data consistent with the distribution of real data through adversarial training of the generator and discriminator (Goodfellow I, Pouget-Abadie J, Mirza M, et al. Generative Adversarial Nets[C] / / Neural Information Processing Systems.MIT Press, 2014.).

[0004] In conclusion, in scenarios where underwater acoustic target sample data is scarce, an underwater acoustic target recognition method that combines acoustic signal feature extraction technology, data augmentation technology, and machine learning methods is essential. Summary of the Invention

[0005] Technical problems to be solved

[0006] To overcome the shortcomings of existing technologies, this invention proposes an underwater acoustic target recognition method based on data augmentation and residual CNN.

[0007] Technical solution

[0008] A method for underwater acoustic target recognition based on data augmentation and residual CNN, characterized by the following steps:

[0009] Step 1: Extract MFCC features of underwater acoustic target signals of each category;

[0010] Step 2: Concatenate the MFCC features of every 5 consecutive frames to generate a two-dimensional matrix. This two-dimensional matrix can be drawn as a color image and used as the input feature image of a single sample, i.e. the initial feature image. Select 3 / 4 of the image samples for each category for model training and validation, and use the remaining 1 / 4 for model testing.

[0011] Step 3: Using traditional data augmentation techniques, set multiple image contrast ranges to obtain generated images with different contrast ranges, and set the scaling and translation ranges of the images in the horizontal and vertical directions.

[0012] Step 4: Input the feature image into the deep convolutional generative adversarial network (DCGAN) model for training. The DCGAN model includes a discriminator and a generator. The two networks work together to improve through adversarial competition in order to achieve the best effect in sample generation, thereby outputting the corresponding generated feature image.

[0013] Step 5: Build a residual CNN classification model based on the ResNet18 model, remove the pooling layers, and adjust its input layer, fully connected layer and output layer to a size suitable for the recognition task. Use this residual CNN model as the classification model for the recognition task, and transform the feature image into a size that matches the size of the convolutional layer in the ResNet18 model.

[0014] Step 6: Use the initial feature image, the multiple generated feature images from Step 3 (after contrast adjustment), and Step 4 (from the deep convolutional generative adversarial network DCGAN model) as input features for the residual CNN model, train the residual CNN model, and then use the test data and data recognition.

[0015] The extraction of underwater acoustic target MFCC features involves: performing an FFT on each frame of signal to obtain the spectrum; filtering the spectrum through a set of triangular bandpass filters to obtain a Mel filter; calculating the logarithmic energy of the output of each filter and its discrete cosine transform to obtain the L-order MFCC; calculating the L MFCC cepstral difference parameters; and combining the three parameters—MFCC, first-order, and second-order cepstral difference parameters—as the feature vector of the signal.

[0016] The initial feature image is processed using traditional data augmentation techniques: the image contrast range is adjusted to 0.1–0.9, 0.2–0.8, and 0.3–0.7, and the original signal image can be converted into three generated images by adjusting the contrast; the scaling range of the image in the horizontal and vertical directions is set to 0.9–1.1, and the translation range is -30 to 30 pixels.

[0017] The Deep Convolutional Adversarial Network (DCGAN) model includes a discriminator and a generator, which are connected. The generator generates simulated samples, and the discriminator judges the authenticity of the samples. The discriminator takes image samples as input and its structure includes four convolutional layers with a kernel size of 5*5, and the number of filters is 64, 128, 256, and 512, respectively. Each convolutional layer is followed by a normalization layer and a ReLU activation layer, and finally a convolutional layer with a kernel size of 4*4, and the number of filters is 1. The generator takes random noise as input and needs to be projected and reshaped. Its structure includes three transposed convolutional layers with a kernel size of 5*5, and the number of filters is 256, 128, and 64, respectively. Each transposed convolutional layer is followed by a normalization layer and a ReLU activation layer, and finally a transposed convolutional layer with a kernel size of 5*5, and the number of filters is 3.

[0018] The residual convolutional neural network ResNet18 model is built by removing the pooling layer from the ResNet18 model and replacing the input layer, fully connected layer and output layer with sizes that match the recognition task. The input layer size is 224*224*1, and the size of the fully connected layer is consistent with the number of target types in the recognition task.

[0019] The image size matches the dimensions of the input layer, fully connected layer, and output layer in the ResNet18 model, which is 224*224.

[0020] Beneficial effects

[0021] This invention proposes an underwater acoustic target recognition method based on data augmentation and residual CNN, which mainly includes extracting MFCC features of underwater acoustic targets, data augmentation processing, and ResNet18 model recognition. It fully leverages the data augmentation advantages of data augmentation technology and the deep feature mining capabilities of residual CNN model to effectively identify underwater acoustic targets.

[0022] Mel-frequency cepstral coefficients (MFCCs) are a widely used feature in speech recognition, proposed based on the characteristics of the human ear. Due to the unique structure of the human ear, listeners can automatically separate the low-frequency and high-frequency components of speech, with the low-frequency component being the main part for identifying speech characteristics.

[0023] The generative adversarial network (GAN) in this study consists of a generator and a discriminator. During training, the generator produces simulated samples, and the discriminator judges the realism of the samples. The basic framework of the GAN is as follows: Figure 1 As shown, the two networks work together to improve through adversarial competition, aiming for optimal sample generation. After training, only the generator network is retained for sample generation. Deep convolutional generative adversarial networks (DCGANs) are derived from GAN models. DCGAN combines CNNs and basic GANs, utilizing deep CNNs for both the generator and discriminator. DCGAN improves the stability and quality of generated results compared to basic GANs.

[0024] Beneficial effects of this invention:

[0025] 1. This study uses the DCGAN model. DCGAN has an excellent image generation architecture. Compared with traditional GAN ​​models, DCGAN training is relatively stable, and the DCGAN discriminator can extract deeper image features by introducing a CNN model, which has a great advantage in image generation and classification.

[0026] 2. This study uses the method of adjusting image contrast, setting three contrast ranges: 0.1 to 0.9, 0.2 to 0.8, and 0.3 to 0.7. Thus, the original feature map can be used to generate three generated maps by adjusting the contrast, thereby expanding the feature image.

[0027] 3. The classification model adopts the ResNet18 model with residual connections. Based on the ResNet18 model, this study removes the pooling layer to retain more underwater acoustic feature information and adjusts the size of the network model to adapt to the feature image input in this study.

[0028] 4. This study used traditional data augmentation techniques and the DCGAN model to expand the feature image obtained by stitching together the MFCC features extracted from underwater acoustic data. Thus, in scenarios where underwater acoustic target sample data is relatively scarce, data augmentation techniques and residual CNN models were used to achieve effective recognition of underwater acoustic targets. Attached Figure Description

[0029] Figure 1 Basic framework of generative adversarial networks

[0030] Figure 2 Overall Flowchart

[0031] Figure 3 Discriminator block diagram

[0032] Figure 4 Generator block diagram

[0033] Figure 5 Deepship dataset: Raw signal waveforms from ships

[0034] Figure 6 MFCC characteristics of different types of ships

[0035] (a) tugboat; (b) cargo ship; (c) oil tanker; (d) passenger ship

[0036] Figure 7 Image samples of different types of ships

[0037] (a) tugboat; (b) cargo ship; (c) oil tanker; (d) passenger ship

[0038] Figure 8 Tugboat image samples under different contrast ranges

[0039] (a)0.1-0.9; (b)0.2-0.8; (c)0.3-0.7

[0040] Figure 9 DCGAN model generates samples from images of different types of ships.

[0041] (a) tugboat; (b) cargo ship; (c) oil tanker; (d) passenger ship Detailed Implementation

[0042] The present invention will now be further described in conjunction with the embodiments and accompanying drawings:

[0043] Mel-frequency cepstral coefficients (MFCCs) are a widely used feature in speech recognition, proposed based on the characteristics of the human ear. Due to the unique structure of the human ear, listeners can automatically separate the low-frequency and high-frequency components of speech, with the low-frequency component being the main part for identifying speech characteristics.

[0044] The traditional data augmentation method used in this study sets the image contrast adjustment range to 0.1–0.9, 0.2–0.8, and 0.3–0.7. The original signal image can be converted into three generated images by adjusting the contrast. In this study, the image scaling range in both the horizontal and vertical directions is set to 0.9–1.1, and the translation range is -30 to 30 pixels.

[0045] Generative Adversarial Networks (GANs) consist of a generator and a discriminator. During training, the generator produces simulated samples, and the discriminator judges the realism of the samples. The basic framework of a GAN is as follows: Figure 1As shown, the two networks work together to improve through adversarial competition, aiming for optimal sample generation. After training, only the generator network is retained for sample generation. Deep convolutional generative adversarial networks (DCGANs) are derived from GAN models. DCGAN combines CNNs and basic GANs, utilizing deep CNNs for both the generator and discriminator. DCGAN improves the stability and quality of generated results compared to basic GANs.

[0046] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0047] 1) This study requires extracting MFCC features from acoustic signals. First, the acquired acoustic signals undergo preprocessing, including pre-emphasis, framing, and windowing. Pre-emphasis enhances the high-frequency components of the signal, making the spectrum smoother. Framing divides the signal into several short time segments, within which the signal can be considered a stationary process. Overlapping segments are typically used during framing to ensure smooth transitions between frames. Windowing reduces signal truncation effects.

[0048] 2) After preprocessing, perform FFT on each frame of signal to obtain the spectrum.

[0049] 3) The spectrum is then filtered through a set of triangular bandpass filters to obtain the Mel filter.

[0050] 4) Calculate the logarithmic energy of the output of each filter and its discrete cosine transform to obtain the Lth order MFCC.

[0051] 5) Calculate the cepstral difference parameter (Delta Cepstrum) based on the L MFCC cepstral coefficient values, and combine the three parts of the parameter (MFCC, first-order and second-order cepstral difference parameter) as the feature vector of the signal.

[0052] 6) Generate an image for every 5 segments of MFCC features, and use this as the feature input for a single sample. Select 3 / 4 of the image samples for each category for model training and validation, and use the remaining 1 / 4 for model testing.

[0053] 7) First, traditional data augmentation techniques are used to adjust the image contrast range and set the horizontal and vertical scaling and translation ranges of the image.

[0054] 8) Build a DCGAN model, input the feature image into the DCGAN model for training, and obtain the corresponding generated image.

[0055] 9) Based on the ResNet18 model, remove the pooling layers and replace the input, fully connected, and output layers with sizes suitable for the recognition task to build a residual CNN model. Then, convert the image size to 224*224 to match the size of the convolutional layers in the ResNet18 model.

[0056] 10) Both initial features and generated features can be used as input features for the neural network, thereby expanding the training samples, training the ResNet18 model, and recognizing the test data.

[0057] Detailed implementation process

[0058] After extracting MFCC features and performing data augmentation, the ResNet18 model achieves sound signal recognition, referring to... Figure 2 The overall process, including setup and training, is divided into the following steps:

[0059] 1) Data preprocessing includes pre-emphasis, framing, and windowing. Pre-emphasis: Boosts the high-frequency components of the signal, making the spectrum smoother. Framing: Divides the signal into several short time intervals, within which the signal can be considered a stationary process. Windowing: Let the signal be s(n), and the window function be w(n). The signal s'(n) obtained after windowing is:

[0060] s'(n)=s(n)w(n) (1)

[0061] Where 0≤n≤N-1, N is the number of sample points, and w(n) is usually taken as a Hamming window.

[0062] 2) After preprocessing, perform FFT on each frame of signal to obtain the spectrum, the discrete spectrum S' of the signal. a (k) is:

[0063]

[0064] 3) The spectrum is filtered through a set of triangular bandpass filters to obtain the Mel filter. There are M filters with center frequencies f(m), where m = 1, 2, ..., M. The formula for the triangular filter is:

[0065]

[0066] 4) Calculate the logarithmic energy of the output of each filter:

[0067]

[0068] 5) Perform a discrete cosine transform on the calculated M logarithmic energies to obtain the L-order MFCC, where L is typically between 12 and 16. The discrete cosine transform formula is:

[0069]

[0070] 6) Calculate the cepstral difference parameters based on the L MFCC cepstral coefficient values, using the following formula:

[0071]

[0072] In the formula, d represents the nth first-order difference result, and C n Let represent the nth cepstral coefficient obtained by equation (5), L represent the order when calculating MFCC, and K represent the time difference of the first derivative, which can be 1 or 2. Substituting the calculation result into equation (6) yields the second-order difference result.

[0073] 7) Combine the three parameters—MFCC, first-order cepstral difference parameter, and second-order cepstral difference parameter—as the signal feature vector input to the GRNN model.

[0074] 8) Generate an image for every 5 segments of MFCC features, and use this as the feature input for a single sample. Select 3 / 4 of the image samples for each category for model training and validation, and use the remaining 1 / 4 for model testing.

[0075] 9) First, traditional data augmentation techniques are used, and the image contrast is adjusted to the ranges of 0.1 to 0.9, 0.2 to 0.8, and 0.3 to 0.7. The original signal image can be converted into three generated images by adjusting the contrast. The scaling range of the image in both the horizontal and vertical directions is set to 0.9 to 1.1, and the translation range is -30 to 30 pixels.

[0076] 10) Build the DCGAN model, whose discriminator and generator frameworks are as follows: Figure 3 and Figure 4 As shown, the feature image is input into the DCGAN model for training, and the corresponding generated image is obtained.

[0077] 11) Based on the ResNet18 model, remove the pooling layers and replace the input, fully connected, and output layers with sizes suitable for the recognition task to build a residual CNN model. Then, convert the image size to 224*224 to match the size of the convolutional layers in the ResNet18 model.

[0078] 12) Both initial features and generated features can be used as input features for the neural network, thereby expanding the training samples, training the ResNet18 model, and recognizing the test data.

[0079] Specific Implementation Examples

[0080] The ship audio data selected in this study comes from the Deepship dataset, which includes four types of vessels: tugboats, cargo ships, tankers, and passenger ships. The initial signals for each type of vessel are as follows: Figure 5 As shown. This study uses MFCC features of audio data as input. First, the audio data is segmented, with 256 samples per segment and 128 samples offset per segment. The MFCC of each segment is then calculated. During the extraction process, 20 filters are set, and their first and second-order difference coefficients are calculated, resulting in a 1*36 feature vector for each segment. This study extracted 6250 MFCC features from various ship types. For tugboats, cargo ships, tankers, and passenger ships, MFCC features from 2.5s audio data are used as examples. Figure 6 As shown.

[0081] Each set of 5 MFCC features generates an image, which is then used as the feature input for a single sample. Figure 7 As shown, 1250 samples were obtained for each type of ship. Three-quarters of the image samples for each category were selected for model training and validation, and the remaining one-quarter were used for model testing. Thus, each type of ship obtained 1000 training samples and 250 test samples.

[0082] The image contrast ranges are set to 0.1–0.9, 0.2–0.8, and 0.3–0.7 respectively. After adjusting the contrast, the original signal image can be converted into three generated images. Taking a tugboat as an example, for instance... Figure 8 As shown. The DCGAN model generates the following results. Figure 9 As shown, the results indicate that after extensive training, the generated results are quite close to the actual feature maps. For an initial signal feature map, this study uses data augmentation methods to obtain a total of four generated feature maps. Both the initial and generated features can be used as input features for the neural network, thereby expanding the training samples. That is, the number of training samples in this study is expanded from 1000 to 5000.

[0083] This study uses Support Vector Machines (SVMs) and standard CNNs as comparative models. For SVMs, images are converted into feature matrices, which are then input into the model for training and to predict classification results. Standard CNNs have a shallower layer count, consisting of only three convolutional layers, denoted as 3_CNN. This study uses recognition accuracy as the metric for the estimation results, as shown in the following formula:

[0084]

[0085] Where N is the sample size. This represents the number of correctly identified samples out of N samples.

[0086] Table 1 shows the recognition accuracy of various methods for ship target signal identification. Each method is compared with and without data augmentation techniques, with Aug representing the case with data augmentation. The results show that, for the same method, using data augmentation significantly improves recognition performance, and the ResNet18 model outperforms both SVM and 3_CNN models. Because deeper CNN models can mine richer data features, the ResNet18 model outperforms the 3_CNN model. Compared to traditional machine learning models, CNN models outperform SVM models due to their advantages in local perception and weight sharing. The ResNet18_Aug model, based on data augmentation and residual CNN, achieves the best recognition results. This is because the ResNet18_Aug model combines the data augmentation advantages of data augmentation with the deep feature mining capabilities of residual CNN models.

[0087] Table 1. Accuracy of Ship Target Signal Recognition by Different Methods

[0088] method tug freighter oil tanker passenger ship Total SVM 78.55％ 78.38％ 79.92％ 77.85％ 78.68％ SVM_Aug 80.29％ 81.31％ 80.96％ 79.32％ 80.47％ 3_CNN 83.18％ 81.91％ 82.35％ 81.03％ 82.12％ 3_CNN_Aug 88.95％ 86.21％ 87.72％ 85.52％ 87.10％ ResNet18 93.11％ 92.24％ 92.98％ 91.62％ 92.48％ ResNet18_Aug 96.85％ 96.23％ 97.01％ 95.38％ 96.37％

Claims

1. A method for underwater acoustic target recognition based on data augmentation and residual CNN, characterized in that... The steps are as follows: Step 1: Extract MFCC features of underwater acoustic target signals of each category; Step 2: Concatenate the MFCC features of every 5 consecutive frames to generate a two-dimensional matrix. This two-dimensional matrix can be drawn as a color image and used as the input feature image of a single sample, i.e. the initial feature image. Select 3 / 4 of the image samples for each category for model training and validation, and use the remaining 1 / 4 for model testing. Step 3: Using traditional data augmentation techniques, multiple image contrast ranges are set to obtain generated images with different contrast ranges, and the scaling and translation ranges in the horizontal and vertical directions of the images are set. Step 4: Input the feature image into the deep convolutional generative adversarial network (DCGAN) model for training. The DCGAN model includes a discriminator and a generator. The two networks work together to improve through adversarial competition in order to achieve the best effect in sample generation, thereby outputting the corresponding generated feature image. Step 5: Build a residual CNN classification model based on the ResNet18 model, remove the pooling layers, and adjust its input layer, fully connected layer and output layer to a size suitable for the recognition task. Use this residual CNN model as the classification model for the recognition task, and transform the feature image into a size that matches the size of the convolutional layer in the ResNet18 model. Step 6: Use the initial feature image, the multiple generated feature images from Step 3 (after contrast adjustment), and Step 4 (from the deep convolutional generative adversarial network DCGAN model) as input features for the residual CNN model, train the residual CNN model, and then use the test data and data recognition.

2. The underwater acoustic target recognition method based on data augmentation and residual CNN according to claim 1, characterized in that: The extraction of underwater acoustic target MFCC features: Perform an FFT on each frame of signal to obtain the spectrum. Filter the spectrum through a set of triangular bandpass filters to obtain the Mel filter. Calculate the logarithmic energy of the output of each filter and its discrete cosine transform to obtain the L-order MFCC. Calculate the L MFCC cepstral difference parameters. Combine the three parameters—MFCC, first-order, and second-order cepstral difference parameters—as the feature vector of the signal.

3. The underwater acoustic target recognition method based on data augmentation and residual CNN according to claim 1, characterized in that: The initial feature image is processed using traditional data augmentation techniques: the image contrast range is adjusted to 0.1–0.9, 0.2–0.8, and 0.3–0.7, and the original signal image can be converted into three generated images by adjusting the contrast; the scaling range of the image in the horizontal and vertical directions is set to 0.9–1.1, and the translation range is -30 to 30 pixels.

4. The underwater acoustic target recognition method based on data augmentation and residual CNN according to claim 1, characterized in that: The Deep Convolutional Adversarial Network (DCGAN) model includes a discriminator and a generator, which are connected. The generator generates simulated samples, and the discriminator judges the authenticity of the samples. The discriminator takes image samples as input and its structure includes four convolutional layers with a kernel size of 5*5, and the number of filters is 64, 128, 256, and 512, respectively. Each convolutional layer is followed by a normalization layer and a ReLU activation layer, and finally a convolutional layer with a kernel size of 4*4, and the number of filters is 1. The generator takes random noise as input and needs to be projected and reshaped. Its structure includes three transposed convolutional layers with a kernel size of 5*5, and the number of filters is 256, 128, and 64, respectively. Each transposed convolutional layer is followed by a normalization layer and a ReLU activation layer, and finally a transposed convolutional layer with a kernel size of 5*5, and the number of filters is 3.

5. The underwater acoustic target recognition method based on data augmentation and residual CNN according to claim 1, characterized in that: The residual convolutional neural network ResNet18 model is built by removing the pooling layer from the ResNet18 model and replacing the input layer, fully connected layer and output layer with sizes that match the recognition task. The input layer size is 224*224*1, and the size of the fully connected layer is consistent with the number of target types in the recognition task.

6. The underwater acoustic target recognition method based on data augmentation and residual CNN according to claim 1, characterized in that: The image size matches the dimensions of the input layer, fully connected layer, and output layer in the ResNet18 model, which is 224*224.

Images