Machine learning method with channel attention pooling
By introducing channel attention pooling into convolutional neural networks, and combining residual networks and pooling layers, the problem of frequent memory access in attention mechanisms is solved, improving computational efficiency and storage space utilization, and enhancing the performance of machine learning models.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NOVATEK MICROELECTRONICS CORP
- Filing Date
- 2025-04-14
- Publication Date
- 2026-06-09
AI Technical Summary
The attention mechanism in convolutional neural networks requires frequent access to dynamic random access memory, which makes the access process time-consuming and consumes a lot of memory space.
By introducing channel attention pooling into convolutional neural networks, the number of accesses to dynamic random access memory is reduced. By combining residual networks and pooling layers with fully connected layers and channel attention networks, efficient processing of feature tensors is achieved.
This reduces the number of accesses to dynamic random access memory, improves computational efficiency and storage space utilization, and enhances the performance of machine learning models.
Smart Images
Figure CN122174887A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a machine learning method, and more particularly to a machine learning method that utilizes channel attention pooling. Background Technology
[0002] Convolutional neural network (CNN) models are commonly used in image processing. In recent years, researchers and companies have begun to adopt attention mechanisms, embedding attention layers into CNN models to achieve better interpretability and performance. The attention mechanism applied in CNN models can learn key features in the input data, generating key feature tensors through channel scaling. The attention mechanism allows the CNN model to adjust its attention level according to different parts of the input, making the model more capable of understanding and interpreting complex data.
[0003] However, attention mechanisms require extensive data access to Dynamic Random Access Memory (DRAM). The processor must access DRAM to load the entire feature tensor, a process that is very time-consuming and consumes a significant amount of memory. Summary of the Invention
[0004] This invention provides a machine learning method using channel attention pooling. The machine learning method includes inputting a first residual input into a convolutional layer of a first residual network to generate a first convolutional output, and inputting the first convolutional output into a first pooling layer to generate a first pooling vector. Attached Figure Description
[0005] Figure 1 This is a schematic diagram of the channel attention method.
[0006] Figure 2 This is a schematic diagram of a machine learning method using channel attention pooling in an embodiment of the present invention.
[0007] Figure 3 This is a schematic diagram of a machine learning method using channel attention pooling in another embodiment of the present invention.
[0008] Figure 4 This is a schematic diagram of a machine learning method using channel attention pooling according to another embodiment of the present invention.
[0009] Figure 5 This is a schematic diagram of a machine learning method using channel attention pooling in another embodiment of the present invention.
[0010] The reference numerals in the attached figures are explained as follows:
[0011] 100: Channel Attention Method
[0012] 101: Input Data
[0013] 102: Convolutional Layer
[0014] 104: Channel Attention Layer
[0015] 106: Feature Tensor of the nth Layer
[0016] 108: Pooling Vector
[0017] 110: Scaling Vector
[0018] 112: Feature tensor of the (n+1)th layer
[0019] 114, 209, 229: Additive layer
[0020] 200, 300, 400, 500: Machine Learning Methods for Channel Attention Pooling
[0021] 202: First Residual Network
[0022] 204: First residual input
[0023] 205, 217: MxM convolutional layers
[0024] 206, 218: Temporary convolution output
[0025] 207, 219: NxN convolutional layers
[0026] 208: Residual Output
[0027] 210: First convolution output
[0028] 212: First pooling layer
[0029] 214: First pooling vector
[0030] 216: Network Input
[0031] 220: First Residual Channel Attention Network
[0032] 221: Channel Attention Network
[0033] 222: First Attention Input
[0034] 223: First Fully Connected Layer
[0035] 224: Temporary fully connected output
[0036] 225: Second Fully Connected Layer
[0037] 226: Fully Connected Output
[0038] 227: Channel-level multiplication layer
[0039] 228: First Attention Output
[0040] 230: Attention Output of the First Residual Channel
[0041] 302: First Residual Network
[0042] 303: First residual input
[0043] 304: Second Residual Network
[0044] 305: Second residual input
[0045] 306: First Residual Channel Attention Network
[0046] 307: Attention Output of the First Residual Channel
[0047] 308: Second Residual Channel Attention Network
[0048] 309: Attention Output of the Second Residual Channel
[0049] 310: Third Residual Channel Attention Network
[0050] 311: Dynamic Random Access Memory
[0051] 312: First pooling layer
[0052] 313: First pooling vector
[0053] 314: The nth residual network
[0054] 315: nth residual input
[0055] 316: The (n-1)th residual input
[0056] 317: Attention Network for the nth Residual Channel
[0057] 318: Attention output of the (n-1)th residual channel
[0058] 319: Attention Output of the nth Residual Channel
[0059] 320: Residual Output
[0060] 321: Network Input
[0061] 322: Attention Network for the (n+1)th Residual Channel
[0062] 401, 501: nth residual input
[0063] 402, 502: First residual network
[0064] 403, 503: First residual input
[0065] 404, 504: Second residual network
[0066] 405, 505: Second residual input
[0067] 406, 506: nth residual network
[0068] 407, 507: The (n-1)th residual input
[0069] 408, 508: First residual channel attention network
[0070] 509: Attention Output of the First Residual Channel
[0071] 409, 511: Attention output of the second residual channel
[0072] 410, 510: Second residual channel attention network
[0073] 411: Attention Output of the Third Residual Channel
[0074] 412, 512: Attention Network for the nth Residual Channel
[0075] 413, 513: Attention output of the nth residual channel
[0076] 414, 514: Dynamic Random Access Memory
[0077] 415, 515: Residual Output
[0078] 416, 516: First pooling layer
[0079] 417: Attention output of the (n+1)th residual channel
[0080] 517: Network Input
[0081] 418, 518: Second pooling layer
[0082] 419, 519: Second convolution output
[0083] 420, 520: The nth pooling layer
[0084] 421, 524: nth pooling vector
[0085] 423, 523: Second pooling vectors
[0086] 424, 521: First pooling vector
[0087] 425, 525: First convolution output
[0088] 426, 526: Output of the nth convolution Detailed Implementation
[0089] Figure 1 This is a schematic diagram of the channel attention method 100. The channel attention method 100 adds a channel attention layer 104 after the convolutional layer 102 of the convolutional neural network (CNN) model. First, the processor inputs the input data 101 (e.g., the input image) into the convolutional layer 102. The convolutional layer 102 convolves the input data 101 and outputs the nth layer feature tensor 106. The nth layer feature tensor 106 is a tensor with dimensions CxHxW, where C is the channel dimension, which is determined by the convolutional layer 102 and contains the features of the input data 101. H is the height dimension and W is the width dimension. The channel attention layer 104 includes a transformation process from the nth layer feature tensor 106 to the (n+1)th layer feature tensor 112. This transformation process involves performing global average pooling (GAP) on the nth layer feature tensor 106 to generate a pooling vector 108. GAP is a process for extracting features from the nth layer feature tensor 106. The GAP process can be calculated as follows:
[0090]
[0091] Among them, u c It is an element of the pooling vector 108, X i,j,c These are elements of the nth layer feature tensor of size 106, where i and j are index values, and each channel of the pooling vector of size 108 contains a u. c .
[0092] The processor then inputs the pooling vector 108 into the fully connected layer and outputs a scaling vector 110. In one embodiment, the number of fully connected layers may be, but is not limited to, two, and the last fully connected layer may be, but is not limited to, sigmoid, ReLU, or softmax. The scaling vector 110 is used to perform channel-level scaling on the nth layer feature tensor 106 to generate the (n+1)th layer feature tensor 112. The nth layer feature tensor 106 is stored in Dynamic Random Access Memory (DRAM) for use during channel-level scaling. The addition layer 114 adds the input data 101 and the (n+1)th layer feature tensor 112 to generate the result. The input data 101 is stored in DRAM for use in the addition layer 114. However, accessing the nth layer feature tensor 106 and the input data 101 (both of which are large in size) is very time-consuming and space-intensive.
[0093] Figure 2This is a schematic diagram of a machine learning method 200 using channel attention pooling in an embodiment of the present invention. First, the processor inputs a first residual input 204 into a first residual network 202 to generate a residual output 208. The first residual input 204 is input data (e.g., a feature tensor) with CxHxW dimensions. The feature tensor can include features of any suitable image or image data.
[0094] The first residual network 202 includes an MxM convolutional layer 205, an NxN convolutional layer 207, and an additive layer 209, where M and N are positive integers. In one embodiment, M is 3 and N is 1. In another embodiment, the MxM convolutional layer 205 and the NxN convolutional layer 207 are used to extract features from the input data. The first residual network 202 may contain, but is not limited to, two convolutional layers. The processor inputs the first residual input 204 into the MxM convolutional layer 205 to generate a temporary convolutional output 206, and inputs the temporary convolutional output 206 into the NxN convolutional layer 207 to generate a first convolutional output 210. The residual output 208 is generated based on the first convolutional output 210 and the first residual input 204. In one embodiment, the residual output 208 is generated by adding the first convolutional output 210 and the first residual input 204 through the additive layer 209. The residual output 208 is a tensor with CxHxW dimensions, i.e. Figure 1 The feature tensor 106 is defined as follows: C is the channel dimension, H is the height dimension, and W is the width dimension. In one embodiment, the residual output 208 is stored in dynamic random access memory (DRAM).
[0095] Next, the processor inputs the first convolutional output 210 into the first pooling layer 212 to generate the first pooling vector 214. The network input 216 is generated based on the residual output 208. The processor inputs the network input 216 into the MxM convolutional layer 217 of the first residual channel attention network 220 to generate the temporary convolutional output 218. Then, the processor inputs the temporary convolutional output 218 into the NxN convolutional layer 219 to generate the first attention input 222. The processor then inputs the first attention input 222 and the first pooling vector 214 into the channel attention network 2 of the first residual channel attention network 220. In step 21, a first attention output 228 is generated. Next, the processor inputs a first pooling vector 214 into a first fully connected layer 223 to generate a temporary fully connected output 224, and then inputs the temporary fully connected output 224 into a second fully connected layer 225 to generate a fully connected output 226. The processor then uses a channel-level multiplication layer 227 to perform channel-level multiplication of the fully connected output 226 with the first attention input 222 to implement the attention mechanism and generate the first attention output 228. Finally, the processor generates a first residual channel attention output 230 based on the first attention output 228 and the network input 216. In this embodiment, the first pooling vector 214 has a small data size and can therefore be stored in the processor instead of in DRAM. In one embodiment, the first pooling layer 212 can be a global average pooling (GAP) layer, a global max pooling layer, a global min pooling layer, an average pooling layer, a max pooling layer, or a min pooling layer. The first pooling vector 214 is a vector with dimension C, where C is the number of channels. Similar to the MxM convolutional layer 205 and the NxN convolutional layer 207, in one embodiment, M is 3 and N is 1. The channel attention network 221 includes multiple fully connected layers and a channel-level multiplication layer 227. In one embodiment, the number of fully connected layers may be, but is not limited to, 2, and the fully connected output 226 is a vector with dimension C used for scaling each channel. In one embodiment, the first residual channel attention output 230 is generated by adding the first attention output 228 and the network input 216 through the addition layer 229. The first residual channel attention output 230 may be an output image (e.g., a deblurred image, a denoised image, and a style-transferred image). In one embodiment, the residual output 208 is stored in dynamic random access memory (DRAM), and the network input 216 is retrieved from DRAM. Figure 2 This method limits the number of DRAM access paths to only one, which is not only efficient but also saves a significant amount of storage space.
[0096] Figure 3This is a schematic diagram of a machine learning method 300 using channel attention pooling in another embodiment of the present invention. The processor inputs the (n-1)th residual channel attention output 318 into the convolutional layer of the nth residual channel attention network 317 to generate the nth attention input, and inputs the first pooling vector 313 and the nth attention input into the channel attention network of the nth residual channel attention network 317 to generate the nth attention output. Furthermore, the nth residual channel attention output 319 is generated based on the (n-1)th residual channel attention output 318 and the nth attention output. Then, the processor inputs the nth residual input 315 into the convolutional layer of the nth residual network 314 to generate the nth convolutional output, and the (n-1)th residual input 316 is generated based on the nth convolutional output and the nth residual input 315. In one embodiment, the first pooling vector 313 is a vector with dimension C, where C is the number of channels. In one embodiment, the nth residual channel attention output 319 is generated by adding the (n-1)th residual channel attention output 318 and the nth attention output through the additive layer of the nth residual channel attention network 317. In another embodiment, the (n-1)th residual input 316 is generated by adding the nth convolutional output and the nth residual input 315 through the additive layer of the nth residual network 314, where the nth residual input 315 is input data (e.g., input image data) with CxHxW dimensions, and n is an integer greater than 1.
[0097] exist Figure 3 In this process, the processor inputs the first residual channel attention output 307 into the second residual channel attention network 308 to generate the second attention input, and inputs the first pooling vector 313 and the second attention input into the channel attention network of the second residual channel attention network 308 to generate the second attention output. The second residual channel attention output 309 is generated based on the first residual channel attention output 307 and the second attention output. Then, the processor inputs the second residual input 305 into the convolutional layer of the second residual network 304 to generate the second convolutional output. The first residual input 303 is generated based on the second convolutional output and the second residual input 305. In one embodiment, the second residual channel attention output 309 is generated by adding the first residual channel attention output and the second attention output through the addition layer of the second residual channel attention network 308. In another embodiment, the first residual input 303 is generated by adding the second convolutional output and the second residual input 305 through the addition layer of the second residual network 304.
[0098] Then, the second residual network 304 outputs the first residual input 303 to the first residual network 302. The first residual network 302 outputs the first convolution output to the first pooling layer 312 to generate the first pooling vector 313. The first residual network 302 stores the residual output 320 in DRAM 311 and accesses the network input 321 from DRAM 311 to the first residual channel attention network 306. The first pooling vector 313 has a small data size, so it can be stored in the processor instead of in DRAM. The first pooling vector 313 can be reused in the first residual channel attention network 306, the second residual channel attention network 308, the third residual channel attention network 310, and the nth residual channel attention network 317. The residual output 320 is a tensor with CxHxW dimensions, i.e. Figure 1 The feature tensor in the model is 106, where C is the channel dimension, H is the height dimension, and W is the width dimension. Figure 3 The only path to access DRAM 311. Therefore, embodiments of the present invention reduce computation time and save storage space by reducing the number of times DRAM 311 is accessed. The first residual channel attention network 306 outputs the first residual channel attention output 307 to the second residual channel attention network 308, the second residual channel attention network 308 outputs the second residual channel attention output 309 to the third residual channel attention network 310, and in one embodiment, the nth residual channel attention network 317 outputs the nth residual channel attention output 319 to the (n+1)th residual channel attention network 322.
[0099] Figure 4This is a schematic diagram of a machine learning method 400 for channel attention pooling according to another embodiment of the present invention. The processor inputs the nth residual input 401 into the convolutional layer of the nth residual network 406 to generate the nth convolutional output 426, and inputs the nth convolutional output 426 into the nth pooling layer 420 to generate the nth pooling vector 421. The (n-1)th residual input 407 is generated based on the nth convolutional output 426 and the nth residual input 401. Then, the processor inputs the (n+1)th residual channel attention output 417 into the convolutional layer of the nth residual channel attention network 412 to generate the nth attention input, and inputs the nth pooling vector 421 and the nth attention input into the channel attention network of the nth residual channel attention network 412 to generate the nth attention output. The nth residual channel attention output 413 is generated based on the (n+1)th residual channel attention output 417 and the nth attention output. In one embodiment, the nth pooling vector 421 is a vector with dimension C, where C is the number of channels. In one embodiment, the nth pooling layer 420 can be a global average pooling (GAP) layer, a global max pooling layer, a global min pooling layer, an average pooling layer, a max pooling layer, or a min pooling layer. The nth residual input 401 is input data (e.g., input image data) with dimension CxHxW. In one embodiment, the (n-1)th residual input 407 is generated by adding the nth convolutional output 426 and the nth residual input 401 through the additive layer of the nth residual network 406. In one embodiment, the nth residual channel attention output 413 is generated by adding the (n+1)th residual channel attention output 417 and the nth attention output through the additive layer of the nth residual channel attention network 412. In one embodiment, the (n+1)th residual channel attention output 417 is the network input 417.
[0100] exist Figure 4In this process, the processor inputs a second residual input 405 into the convolutional layer of the second residual network 404 to generate a second convolutional output 419, and inputs the second convolutional output 419 into the second pooling layer 418 to generate a second pooling vector 423. The first residual input 403 is generated based on the second convolutional output 419 and the second residual input 405. Then, the processor inputs a third residual channel attention output 411 into the convolutional layer of the second residual channel attention network 410 to generate a second attention input, and inputs the second pooling vector 423 and the second attention input into the channel attention network of the second residual channel attention network 410 to generate a second attention output. The second residual channel attention output 409 is generated based on the third residual channel attention output 411 and the second attention output. In one embodiment, the second pooling layer 418 may be a global average pooling (GAP) layer, a global max pooling layer, a global min pooling layer, an average pooling layer, a max pooling layer, or a min pooling layer. In one embodiment, the first residual input 403 is generated by adding the second convolutional output 419 and the second residual input 405 through the addition layer of the second residual network 404. In another embodiment, the second residual channel attention output 409 is generated by adding the third residual channel attention output 411 and the second attention output through the addition layer of the second residual channel attention network 410.
[0101] In addition, the processor inputs the first residual input 403 into the first residual network 402 to generate the first convolution output 425 and the residual output 415 stored in the DRAM 414, and accesses the network input 417 from the DRAM 414 to input into the nth residual channel attention network 412. Furthermore, the processor inputs the first convolution output 425 into the first pooling layer 416 to generate the first pooling vector 424. Then, the processor inputs the first pooling vector 424 into the first residual channel attention network 408, inputs the second residual input 405 into the second residual network 404 to generate the second convolution output 419 and the first residual input 403, and inputs the second convolution output into the second pooling layer 418 to generate the second pooling vector 423. Then, the processor inputs the second pooling vector 423 into the second residual channel attention network 410 to input the nth pooling vector 421 into the nth residual channel attention network 412. The residual output 415 is a tensor with dimensions C x H x W, i.e. Figure 1 The feature tensor is 106, where C is the channel dimension, H is the height dimension, and W is the width dimension. In one embodiment, the pooling vectors 421, 423, and 424 are arranged in a first-in-first-out order. In this embodiment, the DRAM 414 is only written to and read once, thereby reducing computation time and storage space.
[0102] Figure 5This is a schematic diagram of a machine learning method 500 using channel attention pooling in another embodiment of the present invention. The processor inputs the nth residual input 501 into the convolutional layer of the nth residual network 506 to generate the nth convolutional output 526, and inputs the nth convolutional output 526 into the nth pooling layer 520 to generate the nth pooling vector 524. The (n-1)th residual input 507 is generated based on the nth convolutional output 526 and the nth residual input 501. The processor also inputs the (n-1)th residual channel attention output 525 into the convolutional layer of the nth residual channel attention network 512 to generate the nth attention input, and then inputs the nth pooling vector 524 and the nth attention input into the channel attention network of the nth residual channel attention network 512 to generate the nth attention output. The nth residual channel attention output 513 is generated based on the (n-1)th residual channel attention output 525 and the nth attention output. The nth residual input 501 is input data (e.g., input image data) with dimensions C x H x W, and the nth pooling vector 524 is a vector with dimension C, where C is the number of channels. In one embodiment, the nth pooling layer 520 may be a global average pooling (GAP) layer, a global max pooling layer, a global min pooling layer, an average pooling layer, a max pooling layer, or a min pooling layer. In one embodiment, the (n-1)th residual input 507 is generated by adding the nth convolutional output 526 and the nth residual input 501 through the additive layer of the nth residual network 506. In one embodiment, the nth residual channel attention output 513 is generated by adding the (n-1)th residual channel attention output 525 and the nth attention output through the additive layer of the nth residual channel attention network 512.
[0103] exist Figure 5In this process, the processor inputs a second residual input 505 into the convolutional layer of the second residual network 504 to generate a second convolutional output 519, and inputs the second convolutional output 519 into the nth pooling layer 518 to generate a second pooling vector 523. The first residual input 503 is generated based on the second convolutional output 519 and the second residual input 505. Then, the processor inputs a first residual channel attention output 509 into the convolutional layer of the second residual channel attention network 510 to generate a second attention input, and inputs the second pooling vector 523 and the second attention input into the channel attention network of the second residual channel attention network 510 to generate a second attention output. The second residual channel attention output 511 is generated based on the first residual channel attention output 509 and the second attention output. In one embodiment, the second pooling layer 518 may be a global average pooling (GAP) layer, a global max pooling layer, a global min pooling layer, an average pooling layer, a max pooling layer, or a min pooling layer. In one embodiment, the first residual input 503 is generated by adding the second convolutional output 519 and the second residual input 505 through the additive layer of the second residual network 504. In another embodiment, the second residual channel attention output 511 is generated by adding the first residual channel attention output 509 and the second attention output through the additive layer of the second residual channel attention network 510.
[0104] Next, the processor inputs the first residual input 503 into the first residual network 502 to generate the first convolution output 525 and the residual output 515 stored in DRAM 514, and accesses the network input 517 from DRAM 514 to input into the first residual channel attention network 508. Then, the processor inputs the first convolution output 525 into the first pooling layer 516 to generate the first pooling vector 521. Then, the first pooling vector 521 is input into the first residual channel attention network 508, and the second residual input 505 is input into the second residual network 504 to generate the second convolution output 519 and the first residual input 503. Then, the processor inputs the second convolution output 519 into the second pooling layer 518 to generate the second pooling vector 523. Finally, the processor inputs the second pooling vector 523 into the second residual channel attention network 510, and inputs the nth pooling vector 524 into the nth residual channel attention network 512. The residual output 515 is a tensor with dimensions C x H x W, i.e. Figure 1 The feature tensor in the model is 106, where C is the channel dimension, H is the height dimension, and W is the width dimension. Pooling vectors 521, 523, and 524 are arranged in a last-in-first-out (LIFO) order. In this embodiment, DRAM 514 is written to and read from only once, thereby reducing computation time and storage space.
[0105] In summary, the embodiments of the present invention modify the architecture of the channel attention pooling machine learning method, which limits the number of write and read operations to DRAM to only one. Therefore, it can reduce computation time and storage space, thereby improving the efficiency and performance of the machine learning model.
[0106] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A machine learning method utilizing channel attention pooling, characterized in that, include: A first residual input is fed into a convolutional layer of a first residual network to produce a first convolutional output; and The first convolution output is input into a first pooling layer to produce a first pooling vector.
2. The machine learning method as described in claim 1, characterized in that, Feeding the first residual input into the convolutional layer of the first residual network to produce the first convolutional output includes: The first residual input is fed into an MxM convolutional layer to produce a temporary convolutional output; and The temporary convolutional output is input into an NxN convolutional layer to generate the first convolutional output; Where M and N are positive integers.
3. The machine learning method as described in claim 1, characterized in that, The first convolution output is input to the first pooling layer to generate the first pooling vector by inputting the first convolution output to a global average pooling layer, a global max pooling layer, a global min pooling layer, an average pooling layer, a max pooling layer, or a min pooling layer to generate the first pooling vector.
4. The machine learning method as described in claim 1, characterized in that, Also includes: A residual output is generated based on the first convolution output and the first residual input; A network input is fed into a convolutional layer of a first residual channel attention network to produce a first attention input, wherein the network input is generated based on the residual output; The first pooling vector and the first attention input are input into a channel attention network of the first residual channel attention network to generate a first attention output; and A first residual channel attention output is generated based on the network input and the first attention output.
5. The machine learning method as described in claim 4, characterized in that, Feeding the network input into the convolutional layer of the first residual channel attention network to generate the first attention input includes: The network input is fed into an MxM convolutional layer to produce a temporary convolutional output; and The temporary convolutional output is input into an NxN convolutional layer to generate the first attention input; Where M and N are positive integers.
6. The machine learning method as described in claim 4, characterized in that, Inputting the first pooling vector and the first attention input into the channel attention network to generate the first attention output includes: The first pooling vector is input into a first fully connected layer to produce a temporary fully connected output; the temporary fully connected output is input into a second fully connected layer to produce a fully connected output; and The first attention input and the fully connected output are fed into a channel scaling layer to produce the first attention output.
7. The machine learning method as described in claim 4, characterized in that, The residual output is generated by adding the first convolution output and the first residual input.
8. The machine learning method as described in claim 4, characterized in that, The first residual channel attention output is generated by adding the network input and the first attention output together.
9. The machine learning method as described in claim 4, characterized in that, Also includes: The residual output is input into a dynamic random access memory; and The network input is output from the dynamic random access memory.
10. The machine learning method as described in claim 1, characterized in that, Also includes: The attention output of the (n-1)th residual channel is input into the convolutional layer of the nth residual channel attention network to produce the nth attention input; The first pooling vector and the nth attention input are input into a channel attention network of the nth residual channel attention network to generate an nth attention output; and Generate an nth residual channel attention output based on the (n-1)th residual channel attention output and the nth attention output; Where n is an integer greater than 1.
11. The machine learning method as described in claim 10, characterized in that, The nth residual channel attention output is generated by adding the (n-1)th residual channel attention output and the nth attention output.
12. The machine learning method as described in claim 10, characterized in that, Also includes: An nth residual input is fed into a convolutional layer of an nth residual network to produce an nth convolutional output. and A (n-1)th residual input is generated based on the nth convolution output and the nth residual input.
13. The machine learning method as described in claim 12, characterized in that, The (n-1)th residual input is generated by adding the nth convolution output and the nth residual input.
14. The machine learning method as described in claim 1, characterized in that, Also includes: An nth residual input is fed into a convolutional layer of an nth residual network to produce an nth convolutional output. The nth convolution output is input into an nth pooling layer to generate an nth pooling vector; and Generate a (n-1)th residual input based on the nth convolution output and the nth residual input; Where n is an integer greater than 1.
15. The machine learning method as described in claim 14, characterized in that, The (n-1)th residual input is generated by adding the nth convolution output and the nth residual input.
16. The machine learning method as described in claim 14, characterized in that, Also includes: The (n+1)th residual channel attention output is input into the convolutional layer of the nth residual channel attention network to produce the nth attention input; The nth pooling vector and the nth attention input are input into a channel attention network of the nth residual channel attention network to generate an nth attention output; and Generate an nth residual channel attention output based on the (n+1)th residual channel attention output and the nth attention output; The network input is the attention output of the (n+1)th residual channel.
17. The machine learning method as described in claim 16, characterized in that, The nth residual channel attention output is generated by adding the (n+1)th residual channel attention output and the nth attention output.
18. The machine learning method as described in claim 16, characterized in that, Also includes: The residual output is input into a dynamic random access memory; and The attention of the (N+1)th residual channel output from the dynamic random access memory is input into the convolutional layer of the Nth residual network; Where N is the total number of pooling layers.
19. The machine learning method as described in claim 14, characterized in that, Also includes: The attention output of the (n-1)th residual channel is input into the convolutional layer of the nth residual channel attention network to produce the nth attention input; The nth pooling vector and the nth attention input are input into a channel attention network of the nth residual channel attention network to generate an nth attention output; and An nth residual channel attention output is generated based on the (n-1)th residual channel attention output and the nth attention output.
20. The machine learning method as described in claim 19, characterized in that, The nth residual channel attention output is generated by adding the (n-1)th residual channel attention output and the nth attention output.