Signal noise reduction model training method, signal noise reduction method and equipment

Abstract

The invention relates to the technical field of signal processing, provides a training method of a signal noise reduction model, a signal noise reduction method and equipment, and can improve the noise reduction effect of the model. The method comprises the following steps: performing step-by-step noise reduction training on a multistage network based on training signal data corresponding to the multistage network to obtain a signal noise reduction model; if the current level obtained by step-by-step noise reduction training is not the first level, obtaining a loss value of the current level network according to training signal data corresponding to the current level network; and after gradient descent is carried out on the current-level network independently based on the loss value, gradient descent is carried out on the current multi-level network which is subjected to gradient descent independently based on the loss value.

Description

Technical Field

[0001] The present application relates to the field of signal processing technology, and in particular to a signal denoising model training method and device thereof, a signal denoising method and device thereof, a computer device, a storage medium and a computer program product. Background Art

[0002] A signal is a physical quantity that carries information. Most signals carry a certain amount of noise, which is considered interference information. Noise can negatively impact signal processing, reducing signal quality and intelligibility.

[0003] Denoising a signal can reduce the noise in the signal, improve the signal-to-noise ratio, and prevent the signal from being overwhelmed by excessive noise. Traditional techniques can perform denoising using trained models, but the denoising effect of these models is sometimes poor. Summary of the Invention

[0004] Based on this, it is necessary to provide a signal denoising model training method and device, a signal denoising method and device, a computer device, a storage medium and a computer program product to address the above technical problems.

[0005] The present application provides a method for training a signal denoising model, the method comprising:

[0006] Based on the training signal data corresponding to the multi-level network, the multi-level network is trained for noise reduction step by step to obtain a signal noise reduction model;

[0007] Wherein, if the current level trained by the step-by-step noise reduction is not the first level, the loss value of the current level network is obtained according to the training signal data corresponding to the current level network;

[0008] After performing gradient descent on the current level network individually based on the loss value, gradient descent is jointly performed on the current multi-level networks that have already undergone individual gradient descent based on the loss value.

[0009] In one embodiment, the training signal data corresponding to the current level network is added with the signal in the noise removed by the previous level network of the current level network.

[0010] In one embodiment, performing gradient descent on the current level network based on the loss value separately includes:

[0011] Based on the loss value, reversely calculate the current level network layer by layer to obtain a first type of gradient value;

[0012] Based on the first type of gradient value, gradient descent is performed on the current level network.

[0013] In one embodiment, performing gradient descent on the current multi-level individual gradient descent networks based on the loss value includes:

[0014] Based on the loss value, reverse calculation is performed on the current multi-level individual gradient descent network one by one to obtain a second type of gradient value;

[0015] Based on the second type of gradient value, gradient descent is performed on the current multi-level individual gradient descent network.

[0016] In one embodiment, the method further comprises:

[0017] According to the noise reduction calibration range corresponding to each level of the network, the training signal data corresponding to each level of the network is obtained; each training signal data includes training input signal data and training target signal data, the noise level carried by the training input signal data is the upper limit value of the noise reduction calibration range corresponding to the corresponding level of the network, and the noise level carried by the training target signal data is the lower limit value of the noise reduction calibration range corresponding to the corresponding level of the network;

[0018] Obtaining a loss value of the current-level network based on the training signal data corresponding to the current-level network includes:

[0019] Inputting the training input signal data corresponding to the current level network into the current level network;

[0020] The loss value of the current-level network is obtained according to the difference in noise level between the output result of the current-level network and the training target signal data corresponding to the current-level network.

[0021] In one embodiment, before inputting the training input signal data corresponding to the current-level network into the current-level network, the method further includes:

[0022] Obtaining noise removed by a previous-stage network of the current-stage network based on training input signal data and training target signal data corresponding to a previous-stage network;

[0023] Performing signal extraction on the noise removed by the previous level network of the current level network;

[0024] The extracted signal is added to the training input signal data corresponding to the current level network.

[0025] In one embodiment, the multi-level networks are networks with the same network structure, or the multi-level networks are networks with different network structures.

[0026] In one embodiment, when the multi-stage network is a network with different network structures, the structural complexity of the previous stage network in two adjacent stages is higher than the structural complexity of the subsequent stage network, and / or the parameter amount of the previous stage network in two adjacent stages is greater than the parameter amount of the subsequent stage network.

[0027] In one embodiment, the signal data includes any one of CT raw data and CT images.

[0028] In one embodiment, when the signal data includes CT raw data, and the CT raw data is CT multi-level data, the training input CT multi-level data is merged and input into the network or is independently input into the network;

[0029] In the case where the signal data includes a CT image, and the CT image is a CT multi-energy level image, the training input CT multi-energy level image is merged and input into the network or is independently input into the network.

[0030] The present application provides a training device for a signal denoising model, the device comprising:

[0031] A model acquisition module is used to perform noise reduction training on the multi-level network step by step based on the training signal data corresponding to the multi-level network to obtain a signal noise reduction model;

[0032] The training module is used to obtain the loss value of the current-level network according to the training signal data corresponding to the current-level network if the current level trained by the step-by-step noise reduction is not the first level; after performing gradient descent on the current-level network based on the loss value, the current multi-level network that has been gradient descended separately is jointly gradient descended based on the loss value.

[0033] The present application provides a signal noise reduction method, the method comprising:

[0034] Obtaining a signal to be denoised;

[0035] The signal to be denoised is input into a signal denoising model to obtain a denoised signal; the signal denoising model is trained according to the signal denoising model training method introduced in the above embodiment.

[0036] In one embodiment, inputting the signal to be denoised into a signal denoising model to obtain a denoised signal includes:

[0037] Determining the noise level of the signal to be denoised;

[0038] Determining a target input network in a signal noise reduction model according to the noise level of the signal to be denoised; the noise level of the signal to be denoised is within a noise reduction calibration range of the target input network;

[0039] The signal to be denoised is input into the target input network, so that the forward multi-stage network starting from the target input network in the signal denoising model performs step-by-step denoising on the signal to be denoised to obtain a denoised signal.

[0040] The present application provides a signal noise reduction device, comprising:

[0041] A signal acquisition module, used to acquire the signal to be denoised;

[0042] The denoising module is used to input the signal to be denoised into a signal denoising model to obtain a denoised signal; the signal denoising model is trained according to the signal denoising model training method introduced in the above embodiment.

[0043] The present application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the above method.

[0044] The present application provides a computer-readable storage medium having a computer program stored thereon, and the computer program is used by a processor to execute the above method.

[0045] The present application provides a computer program product having a computer program stored thereon, wherein the computer program is used by a processor to execute the above method.

[0046] In the solution provided by the present application, based on the training signal data corresponding to the multi-level network, the multi-level network is trained for step-by-step noise reduction to obtain a signal noise reduction model; wherein, if the current level trained for step-by-step noise reduction is not the first level, the loss value of the current level network is obtained based on the training signal data corresponding to the current level network; after gradient descent is performed on the current level network separately based on the loss value, the current multi-level networks that have already undergone separate gradient descent are jointly gradient descended based on the loss value. The present application adopts a step-by-step noise reduction method, where each level of the network reduces some noise; during the training process, not only a separate gradient descent is performed on the single-level network, but also a joint gradient descent is performed on the multi-level network, which can improve the noise reduction ability of the signal noise reduction model and enhance the noise reduction effect of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0048] Figure 1 1 is a flow chart of a method for training a signal denoising model in one embodiment;

[0049] FIG2( a ) is a schematic diagram of a series connection of a four-level network in one embodiment;

[0050] FIG2( b ) is a schematic diagram of the architecture of the first-stage noise reduction training in one embodiment;

[0051] FIG2( c ) is a schematic diagram of the architecture of the second-stage noise reduction training in one embodiment;

[0052] FIG2( d ) is a schematic diagram of the architecture of the third-level noise reduction training in one embodiment;

[0053] Figure 2(e) is a schematic diagram of the architecture of the fourth-level noise reduction training in one embodiment;

[0054] FIG3( a ) is a schematic diagram of the architecture of first-stage noise reduction training for CT multi-energy level data in one embodiment;

[0055] FIG3( b ) is a schematic diagram of the architecture of the second-stage noise reduction training for CT multi-energy level data in one embodiment;

[0056] Figure 4 1 is a flow chart of a signal noise reduction method according to an embodiment;

[0057] Figure 5 1 is a flow chart of the reasoning phase in one embodiment;

[0058] Figure 6 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0060] The signal denoising model training method and signal denoising method provided in this application can be executed by a computer device. The signal denoising model training method may include Figure 1 Steps shown.

[0061] Step S101: Based on the training signal data corresponding to the multi-stage network, the multi-stage network is subjected to step-by-step noise reduction training to obtain a signal noise reduction model.

[0062] When constructing a signal noise reduction model, the number of networks to be included in the signal noise reduction model can be determined according to actual needs, and the number of networks can be set to be greater than or equal to 2.

[0063] Multiple networks can be connected in series. Accordingly, there is a hierarchical relationship between networks: the closer the network is to the model input, the lower its level. The network closest to the model input is called the first-level network, and the network closest to the model output is called the last-level network. Let n be the integer number of networks. For example, let n be 4. As shown in Figure 2(a), in a four-level network, the first-level network is closest to the model input, and the fourth-level network is closest to the model output.

[0064] Calibrate the corresponding noise reduction range for each network level. Networks closer to the model input (i.e., networks with smaller levels) have larger noise reduction ranges, while networks closer to the output (i.e., networks with larger levels) have smaller noise reduction ranges.

[0065] In the order of network levels from large to small, the noise reduction calibration range corresponding to the multi-level network decreases successively, showing a decreasing relationship; moreover, the noise reduction calibration ranges corresponding to the multi-level network can be independent of each other, and the noise reduction calibration ranges corresponding to the multi-level network can together constitute a larger range.

[0066] The noise reduction calibration range can be divided by absolute noise level or relative noise level. Absolute noise level is the actual physical measurement of noise and is a direct measure of the noise level itself. Relative noise level is a numerical comparison of the noise level relative to a specific noise level and can be expressed as a percentage.

[0067] Let n be the number of networks, and for this example, n = 4. If the noise reduction calibration range is divided by absolute noise level, then the first-level network can have a range of 400 to 300, the second-level network can have a range of 300 to 200, the third-level network can have a range of 200 to 100, and the fourth-level network can have a range of 100 to 0. The noise reduction calibration ranges for these four levels decrease in descending order of network level. These ranges can be independent of each other and can collectively form a larger range of 400 to 0.

[0068] If the noise reduction calibration range is divided by relative noise level, the noise reduction calibration range corresponding to the first-level network can be 100% to 75%, the noise reduction calibration range corresponding to the second-level network can be 75% to 50%, the noise reduction calibration range corresponding to the third-level network can be 50% to 25%, and the noise reduction calibration range corresponding to the fourth-level network can be 25% to 0. The noise reduction calibration ranges corresponding to these four-level networks decrease in order from network level to network level. The noise reduction calibration ranges can be independent of each other and can together form a larger range of 100% to 0.

[0069] The spans of the noise reduction calibration ranges for multiple-stage networks can be the same or different. The span of the noise reduction calibration range for a single-stage network is calculated by subtracting the lower limit of the noise reduction calibration range from the upper limit of the noise reduction calibration range for that stage of the network.

[0070] When the noise reduction calibration ranges for multiple networks are of the same span, the noise reduction calibration ranges for the multiple networks decrease proportionally. When the noise reduction calibration ranges for the multiple networks are of different spans, the noise reduction calibration ranges for the multiple networks can be reduced according to a logarithmic function.

[0071] When the noise reduction calibration range is divided by absolute noise level, in some scenarios, the noise reduction calibration range corresponding to the first-level network is 400 to 300, the noise reduction calibration range corresponding to the second-level network is 300 to 200, the noise reduction calibration range corresponding to the third-level network is 200 to 100, and the noise reduction calibration range corresponding to the fourth-level network is 100 to 0. In this scenario, the noise reduction calibration range corresponding to multiple levels of the network has the same span of 100.

[0072] When the noise reduction calibration range is divided by absolute noise level, in other scenarios, the noise reduction calibration range corresponding to the first-level network is 400 to 350, the noise reduction calibration range corresponding to the second-level network is 350 to 250, the noise reduction calibration range corresponding to the third-level network is 250 to 100, and the noise reduction calibration range corresponding to the fourth-level network is 100 to 0. In this scenario, the noise reduction calibration ranges corresponding to the multi-level networks have different spans. The span of the noise reduction calibration range corresponding to the first-level network is 50, the span of the noise reduction calibration range corresponding to the second-level network is 100, the span of the noise reduction calibration range corresponding to the third-level network is 150, and the span of the noise reduction calibration range corresponding to the fourth-level network is 100.

[0073] When noise reduction calibration ranges are divided by relative noise levels, in some scenarios, the first-stage network has a range of 100% to 75%, the second-stage network has a range of 75% to 50%, the third-stage network has a range of 50% to 25%, and the fourth-stage network has a range of 25% to 0. In this scenario, the noise reduction calibration ranges for multiple stages of the network have the same span of 25%.

[0074] When the noise reduction calibration range is divided by relative noise level, in other scenarios, the noise reduction calibration range corresponding to the first-stage network is 100% to 85%, the noise reduction calibration range corresponding to the second-stage network is 85% to 60%, the noise reduction calibration range corresponding to the third-stage network is 60% to 25%, and the noise reduction calibration range corresponding to the fourth-stage network is 25% to 0. In this scenario, the noise reduction calibration ranges corresponding to the multiple-stage networks have different spans. The span of the noise reduction calibration range corresponding to the first-stage network is 15%, the span of the noise reduction calibration range corresponding to the second-stage network is 25%, the span of the noise reduction calibration range corresponding to the third-stage network is 35%, and the span of the noise reduction calibration range corresponding to the fourth-stage network is 25%.

[0075] The noise reduction calibration range corresponding to each level of the network is used to calibrate that each level of the network can reduce the signal-carrying noise from a certain size to another size. The noise size before reduction corresponds to the upper limit value of the noise reduction calibration range, and the noise size before and after reduction corresponds to the lower limit value of the noise reduction calibration range.

[0076] Therefore, after determining the noise reduction calibration range corresponding to each level of the network, the noise level carried by the training signal data input to each level of the network and the noise level to be reduced can be determined based on the noise reduction calibration range of each level of the network itself; based on the noise level carried by the training signal data input to each level of the network, training signal data with corresponding noise level as input (which can be called training input signal data) is constructed; based on the noise level to be reduced, training signal data with corresponding noise level as target (which can be called training target signal data) is constructed; the training input signal data and the training target signal data form a training signal data.

[0077] It can be understood that each level of the network may correspond to at least one training signal data.

[0078] After obtaining the training signal data corresponding to the multi-level network, the multi-level network can be trained for noise reduction step by step based on the training signal data. The training is carried out in order of network level from smallest to largest. The first level network is trained first, then the second level network, and so on, and finally the last level network is trained. After all levels of the network are trained, one iteration is considered complete. Multiple iterations can be performed according to actual needs.

[0079] The following describes the relevant training process of a single iteration.

[0080] If the current level in the progressive noise reduction training is the first level, that is, the first-level noise reduction training is currently in progress, the loss value of the first-level network can be obtained based on the training signal data corresponding to the first-level network. Based on the loss value of the first-level network, gradient descent is performed on the first-level network alone. At this point, the first-level noise reduction training is completed, and the next-level noise reduction training can be carried out.

[0081] Step S102, where if the current level trained by the step-by-step noise reduction is not the first level, the loss value of the current level network is obtained according to the training signal data corresponding to the current level network; after the current level network is gradient descended separately based on the loss value, the current multi-level gradient descended networks are jointly gradient descended based on the loss value.

[0082] For example, if the current level of the step-by-step noise reduction training is the second level, that is, the second-level noise reduction training is currently being performed, the loss value of the second-level network can be obtained based on the training signal data corresponding to the second-level network. Based on the loss value of the second-level network, gradient descent is performed on the second-level network separately. After gradient descent is performed on the second-level network separately, the current multi-level network that has been gradient-descented separately includes the second-level network and a network lower than the second level (i.e., the first-level network). Based on the loss value of the second-level network, gradient descent can be performed jointly on the first-level network and the second-level network. At this point, the second-level noise reduction training is completed, and the next-level noise reduction training can be performed.

[0083] The current level of the step-by-step noise reduction training is the third level. This means that the third-level noise reduction training is currently in progress. Based on the training signal data corresponding to the third-level network, the loss value of the third-level network can be obtained. Based on the loss value of the third-level network, gradient descent is performed on the third-level network separately. After gradient descent is performed on the third-level network separately, the current multi-level network that has undergone gradient descent includes the third-level network and the networks below the third level (i.e., the first-level network and the second-level network). Based on the loss value of the third-level network, gradient descent can be performed jointly on the first-level network, the second-level network, and the third-level network. At this point, the third-level noise reduction training is complete, and the next-level noise reduction training can be performed.

[0084] The remaining levels of denoising training are performed in the same manner as described above until the final level of denoising training is completed. After the final level of denoising training is completed, one iteration can be considered completed.

[0085] In the training method of the above-mentioned signal denoising model, based on the training signal data corresponding to the multi-level network, the multi-level network is trained for denoising step by step to obtain a signal denoising model; wherein, if the current level trained by the step-by-step denoising is not the first level, the loss value of the current level network is obtained according to the training signal data corresponding to the current level network; after gradient descent is performed on the current level network separately based on the loss value, the current multi-level networks that have been gradient descended separately are jointly gradient descended based on the loss value. The present application adopts a step-by-step denoising method, and each level of the network reduces part of the noise; during the training process, not only the single-level network is gradient descended separately, but also the multi-level network is jointly gradient descended, which can improve the denoising ability of the signal denoising model and enhance the denoising effect of the model.

[0086] In one embodiment, the training signal data corresponding to the current level network is added with the signal in the noise removed by the previous level network of the current level network.

[0087] For example, if the current level of the progressive noise reduction training is the first level, that is, the first-level noise reduction training is currently in progress, the loss value of the first-level network can be obtained based on the training signal data corresponding to the first-level network. Based on the loss value of the first-level network, gradient descent is performed on the first-level network alone. At this point, the first-level noise reduction training is completed, and the next-level noise reduction training can be carried out.

[0088] The current level of step-by-step noise reduction training is the second level. That is, during the second-level noise reduction training, the signal from the noise removed by the first-level network can be added to the training signal data corresponding to the second-level network. The loss value of the second-level network can be obtained based on the training signal data corresponding to the second-level network with this signal added. Based on the loss value of the second-level network, gradient descent is performed on the second-level network alone. After gradient descent is performed on the second-level network alone, the current multi-level network, which has been gradient-descented individually, now includes the second-level network and a network below the second level (i.e., the first-level network). Based on the loss value of the second-level network, gradient descent can be performed jointly on the first-level and second-level networks. At this point, the second-level noise reduction training is complete, and the next-level noise reduction training can be performed.

[0089] The current level of the step-by-step noise reduction training is the third level. That is, during the third-level noise reduction training, the signal from the noise removed by the second-level network can be added to the training signal data corresponding to the third-level network. The loss value of the third-level network can be obtained based on the training signal data corresponding to the third-level network with the added signal. Based on the loss value of the third-level network, gradient descent is performed on the third-level network separately. After gradient descent is performed on the third-level network separately, the current multi-level network, which has been gradient-descented separately, now includes the third-level network and the networks below the third level (i.e., the first-level network and the second-level network). Based on the loss value of the third-level network, gradient descent can be performed jointly on the first-level network, the second-level network, and the third-level network. At this point, the third-level noise reduction training is complete, and the next-level noise reduction training can be performed.

[0090] The remaining levels of denoising training are performed in the same manner as described above until the final level of denoising training is completed. After the final level of denoising training is completed, one iteration can be considered completed.

[0091] Specifically, when adding the signal from the noise removed by the previous network to the training signal data corresponding to the current network, a set proportion of the signal from the noise removed by the previous network is added to the training signal data corresponding to the current network. The set proportion can be set based on actual needs and can be between 0 and 1. The set proportions used in different levels of noise reduction training can be the same or different.

[0092] In a scenario where the same set ratio is used in noise reduction training at different levels, for example, 80% of the signal in the noise removed by the first-level network is added to the training signal data corresponding to the second-level network; and 80% of the signal in the noise removed by the second-level network is added to the training signal data corresponding to the third-level network.

[0093] In scenarios where different ratios are set for different levels of noise reduction training, for example, 90% of the signals in the noise removed by the first-level network are added to the training signal data corresponding to the second-level network; and 80% of the signals in the noise removed by the second-level network are added to the training signal data corresponding to the third-level network.

[0094] In the process of step-by-step noise reduction training, this embodiment adds the signal in the noise removed by the previous level network to the training signal data corresponding to the current level network to compensate and restore the signal, thereby avoiding the situation where the signal is weakened after noise reduction at each level network during the training process. This allows the signal noise reduction model to perform noise reduction while ensuring signal accuracy. In particular, when the signal noise reduction model processes some signals carrying large noise, the situation where the signal is weakened can be significantly alleviated.

[0095] In one embodiment, gradient descent is performed on the current level network based on the loss value, including:

[0096] Based on the loss value, the current level network is reversely calculated layer by layer to obtain the first type of gradient value; based on the first type of gradient value, the current level network is gradient descended.

[0097] Each network level can include multiple layers. Therefore, when performing gradient descent on a single-level network, we can perform a layer-by-layer reverse calculation on the network to obtain the gradient value. The direction of the layer-by-layer reverse calculation is from the layer closest to the model output to the layer closest to the model input. The gradient value obtained by the layer-by-layer reverse calculation is mainly used for independent gradient descent. To distinguish them, the gradient value obtained by the layer-by-layer reverse calculation can be called the first-class gradient value.

[0098] This article uses the current first-level noise reduction training as an example. Based on the training signal data corresponding to the first-level network, the loss value of the first-level network is obtained. Based on the loss value of the first-level network, the first-level network is reversely calculated layer by layer to obtain the first-class gradient value. Based on the first-class gradient value, the first-level network is gradient descent to update the parameters within the first-level network.

[0099] This embodiment performs a layer-by-layer reverse calculation on the current-level network based on the loss value of the current-level network to obtain a first-class gradient value. Based on the first-class gradient value, the current-level network is gradient-decreased so that the parameters in the current-level network can be updated in the direction of decreasing the loss function value.

[0100] In one embodiment, the current multi-level gradient descent networks that have been individually gradient descented are jointly gradient descented based on the loss value, including:

[0101] Based on the loss value, the current multi-level individual gradient descent network is reversely calculated network by network to obtain the second type of gradient value; based on the second type of gradient value, the current multi-level individual gradient descent network is gradient descended.

[0102] During joint gradient descent, if the number of levels in the current individual gradient descent network is greater than or equal to two, a network-by-network reverse calculation is performed on the current multi-level individual gradient descent network to obtain the gradient value. This network-by-network reverse calculation corresponds to the direction from the network closest to the model output to the network closest to the model input. The gradient value obtained by the network-by-network reverse calculation is mainly used in joint gradient descent. To distinguish them, the gradient value obtained by the network-by-network reverse calculation can be referred to as the second type of gradient value.

[0103] For example, consider the current second-level noise reduction training. The current multi-level, individually gradient-descent network includes the first-level network and the second-level network. Based on the loss value of the second-level network, a reverse calculation is performed on each of the first-level and second-level networks to obtain the second-type gradient value. Based on this second-type gradient value, gradient descent is performed on both the first-level and second-level networks.

[0104] For example, consider the current third-level noise reduction training. The current multi-level, individually gradient-descent network includes the first-level network, the second-level network, and the third-level network. Based on the loss value of the third-level network, a reverse calculation is performed on each of the first, second, and third-level networks to obtain the second-type gradient value. Based on the second-type gradient value, gradient descent is performed on the first, second, and third-level networks.

[0105] This embodiment performs a layer-by-layer reverse calculation of the current multi-level, independently gradient-descent network based on the loss value of the current level network to obtain a second type of gradient value. Based on the second type of gradient value, the current multi-level, independently gradient-descent network is gradient-descented, so that the parameters within the current multi-level, independently gradient-descent network can be updated in a direction that reduces the loss function value. This joint gradient descent approach can converge to the optimal solution more quickly.

[0106] In one embodiment, the training method of the signal denoising model provided in the present application further includes:

[0107] According to the noise reduction calibration range corresponding to each level of the network, the training signal data corresponding to each level of the network is obtained; each training signal data includes training input signal data and training target signal data, the noise size carried by the training input signal data is the upper limit value of the noise reduction calibration range corresponding to the corresponding level of the network, and the noise size carried by the training target signal data is the lower limit value of the noise reduction calibration range corresponding to the corresponding level of the network.

[0108] According to the training signal data corresponding to the current level network, the loss value of the current level network is obtained, including:

[0109] The training input signal data corresponding to the current level network is input into the current level network; the loss value of the current level network is obtained according to the difference in noise size between the output result of the current level network and the training target signal data corresponding to the current level network.

[0110] The noise reduction calibration range corresponding to each level of the network is used to calibrate that each level of the network can reduce the signal-carrying noise from a certain size to another size. The noise size before reduction corresponds to the upper limit value of the noise reduction calibration range, and the noise size before and after reduction corresponds to the lower limit value of the noise reduction calibration range.

[0111] Therefore, after determining the noise reduction calibration range corresponding to each level of the network, the noise level carried by the training signal data input to each level of the network can be determined based on the noise reduction calibration range of each level of the network itself, and then the training input signal data of the corresponding noise level can be constructed; according to the noise reduction calibration range of each level of the network itself, the noise level to be reduced can be determined, and then the training target signal data of the corresponding noise level can be constructed; the training input signal data and the training target signal data form a training signal data.

[0112] In scenarios where the noise reduction calibration range is divided by absolute noise level, this article uses a first-stage network with a noise reduction calibration range of 400 to 300 as an example. The upper limit of the noise reduction calibration range for the first-stage network is 400, which allows the generation of training input signal data with a noise level of 400. The lower limit of the noise reduction calibration range for the first-stage network is 300, which allows the generation of training target signal data with a noise level of 300.

[0113] In some scenarios, a preset noise-free signal can be obtained; for any level of network, noise with a noise size of the upper limit of the noise reduction calibration range corresponding to the network at that level is added to the noise-free signal to obtain the training input signal data corresponding to the network at that level, and noise with a noise size of the lower limit of the noise reduction calibration range corresponding to the network at that level is added to the noise-free signal to obtain the training target signal data corresponding to the network at that level.

[0114] This example uses a first-stage network with a noise reduction calibration range of 400 to 300 as an example. After obtaining a noise-free signal, we can simulate the generation of noise with a magnitude of 400 and add this noise to the noise-free signal, thereby obtaining training input signal data with an absolute noise magnitude of 400. We can also simulate the generation of noise with a magnitude of 300 and add this noise to the noise-free signal, thereby obtaining training target signal data with an absolute noise magnitude of 300.

[0115] In other scenarios, a preset noisy signal can be obtained and the noise level of the noisy signal can be determined; for any level of network, the relative size relationship between the noise level of the noisy signal and the upper limit value of the noise reduction calibration range can be determined, and the relative size relationship between the noise level of the noisy signal and the lower limit value of the noise reduction calibration range can be determined.

[0116] If the noise of the noisy signal is greater than the upper limit of the noise reduction calibration range, and the noise of the noisy signal is greater than the lower limit of the noise reduction calibration range, the noise size of the noisy signal can be reduced to the upper limit of the noise reduction calibration range through the preset noise reduction algorithm, and then the training input signal data corresponding to the network at this level is obtained; the noise size of the noisy signal can be reduced to the lower limit of the noise reduction calibration range through the noise reduction algorithm, and then the training target signal data corresponding to the network at this level is obtained.

[0117] If the noise of the noisy signal is less than the upper limit of the noise reduction calibration range, and the noise of the noisy signal is greater than the lower limit of the noise reduction calibration range, then noise can be added to the noisy signal so that the noise level of the noisy signal increases to the upper limit of the noise reduction calibration range, thereby obtaining the training input signal data corresponding to the network at this level; through the noise reduction algorithm, the noise level of the noisy signal is reduced to the lower limit of the noise reduction calibration range, thereby obtaining the training target signal data corresponding to the network at this level.

[0118] If the noise of the noisy signal is less than the upper limit of the noise reduction calibration range, and the noise of the noisy signal is less than the lower limit of the noise reduction calibration range, then noise can be added to the noisy signal so that the noise level of the noisy signal increases to the upper limit of the noise reduction calibration range, thereby obtaining the training input signal data corresponding to the network at this level; and noise can be added to the noisy signal so that the noise level of the noisy signal increases to the lower limit of the noise reduction calibration range, thereby obtaining the training target signal data corresponding to the network at this level.

[0119] If the noise of the noisy signal is equal to the upper limit of the noise reduction calibration range, the noisy signal can be directly used as the training input signal data corresponding to the network at this level.

[0120] If the noise of the noisy signal is equal to the lower limit of the noise reduction calibration range, the noisy signal can be directly used as the training target signal data corresponding to the network at this level.

[0121] The preset noise reduction algorithms may include but are not limited to low-pass filtering, iterative noise reduction, deep learning noise reduction and other algorithms.

[0122] The first-level network with a noise reduction calibration range of 400 to 300 is used as an example for introduction. According to the method described above, the training input signal data with an absolute noise size of 400 corresponding to the first-level network can be obtained, and the training target signal data with an absolute noise size of 300 corresponding to the first-level network can be obtained. Then, the training input signal data with an absolute noise size of 400 corresponding to the first-level network can be input into the first-level network to obtain the output result of the first-level network; based on the difference in noise size between the output result of the first-level network and the training target signal data with an absolute noise size of 300 corresponding to the first-level network, the loss value of the first-level network can be obtained. Subsequently, separate gradient descent and joint gradient descent can be performed based on the loss value of the first-level network.

[0123] This article uses a second-level network with a noise reduction calibration range of 300 to 200 as an example. Following the aforementioned approach, we can obtain training input signal data with an absolute noise level of 300 for the second-level network, as well as training target signal data with an absolute noise level of 200 for the second-level network. This training input signal data with an absolute noise level of 300 for the second-level network can then be input into the second-level network to obtain the output of the second-level network. Based on the difference in noise level between the output of the second-level network and the training target signal data with an absolute noise level of 200 for the second-level network, the loss value of the second-level network can be determined. Subsequently, separate and joint gradient descent can be performed based on the loss value of the second-level network.

[0124] In scenarios where the noise reduction calibration range is divided by relative noise level, we'll use a first-stage network with a noise reduction calibration range of 100% to 75% as an example. In this scenario, you can first calibrate the noise level, which is called the calibrated noise value. For example, a calibrated noise value of 500 is used.

[0125] Since the upper limit of the noise reduction calibration range corresponding to the first-level network is 100%, the product of the calibrated noise value and the upper limit is 500, and the training input signal data with a noise level of 500 can be constructed. Since the lower limit of the noise reduction calibration range corresponding to the first-level network is 75%, the product of the calibrated noise value and the lower limit is 375, and the training target signal data with a noise level of 375 can be constructed.

[0126] In some scenarios, a preset noise-free signal can be obtained. For any level of the network, the product of the upper limit of the noise reduction calibration range corresponding to that level of the network and the calibrated noise value (recorded as the first multiplication result) and the product of the lower limit of the noise reduction calibration range corresponding to that level of the network and the calibrated noise value (recorded as the second multiplication result) can be obtained. Noise corresponding to the first multiplication result is added to the noise-free signal to obtain the training input signal data for that level of the network. Noise corresponding to the second multiplication result is added to the noise-free signal to obtain the training target signal data for that level of the network.

[0127] This example uses a first-stage network with a noise reduction calibration range of 100% to 75% and a calibrated noise value of 500. The first multiplication result of the upper limit of the noise reduction calibration range corresponding to the first-stage network and the calibrated noise value is 500, and the first multiplication result of the lower limit of the noise reduction calibration range corresponding to the first-stage network and the calibrated noise value is 375. After obtaining a noise-free signal, you can simulate the generation of noise with a magnitude of 500 and add this noise to the noise-free signal to obtain training input signal data with a noise magnitude of 500. You can also simulate the generation of noise with a magnitude of 375 and add this noise to the noise-free signal to obtain training target signal data with a noise magnitude of 375.

[0128] In other scenarios, a preset noisy signal can be obtained and the noise level of the noisy signal can be determined; for any level of network, the relative size relationship between the noise level of the noisy signal and the first multiplication result can be determined, and the relative size relationship between the noise level of the noisy signal and the second multiplication result can be determined.

[0129] If the noise of the noisy signal is greater than the first multiplication result, and the noise of the noisy signal is greater than the second multiplication result, the preset noise reduction algorithm can be used to reduce the noise size of the noisy signal to the numerical value corresponding to the first multiplication result, thereby obtaining the training input signal data corresponding to the network at this level; the noise reduction algorithm can be used to reduce the noise size of the noisy signal to the numerical value corresponding to the second multiplication result, thereby obtaining the training target signal data corresponding to the network at this level.

[0130] If the noise of the noisy signal is smaller than the first multiplication result, and the noise of the noisy signal is larger than the second multiplication result, noise can be added to the noisy signal so that the noise size of the noisy signal increases to the numerical value corresponding to the first multiplication result, thereby obtaining the training input signal data corresponding to the network at this level; through the noise reduction algorithm, the noise size of the noisy signal is reduced to the numerical value corresponding to the second multiplication result, thereby obtaining the training target signal data corresponding to the network at this level.

[0131] If the noise of the noisy signal is smaller than the first multiplication result, and the noise of the noisy signal is smaller than the second multiplication result, noise can be added to the noisy signal so that the noise size of the noisy signal increases to the numerical value corresponding to the first multiplication result, thereby obtaining the training input signal data corresponding to the network at this level; noise can be added to the noisy signal so that the noise size of the noisy signal increases to the numerical value corresponding to the second multiplication result, thereby obtaining the training target signal data corresponding to the network at this level.

[0132] If the noise of the noisy signal is equal to the first multiplication result, the noisy signal can be directly used as the training input signal data corresponding to the network at this level.

[0133] If the noise of the noisy signal is equal to the second multiplication result, the noisy signal can be directly used as the training target signal data corresponding to the network at this level.

[0134] The preset noise reduction algorithms may include but are not limited to low-pass filtering, iterative noise reduction, deep learning noise reduction and other algorithms.

[0135] The first-level network with a noise reduction calibration range of 100% to 75% is used as an example for introduction. According to the method described above, the training input signal data with a relative noise level of 100% corresponding to the first-level network can be obtained, and the training target signal data with a relative noise level of 75% corresponding to the first-level network can be obtained. Then, the training input signal data with a relative noise level of 100% corresponding to the first-level network can be input into the first-level network to obtain the output result of the first-level network; based on the difference in noise level between the output result of the first-level network and the training target signal data with a relative noise level of 75% corresponding to the first-level network, the loss value of the first-level network can be obtained. Subsequently, separate gradient descent and joint gradient descent can be performed based on the loss value of the first-level network.

[0136] This article uses a second-stage network with a noise reduction calibration range of 75% to 50% as an example. Following the aforementioned method, we can obtain training input signal data with a relative noise level of 75% for the second-stage network, and obtain training target signal data with a relative noise level of 50% for the second-stage network. Then, we can input the training input signal data with a relative noise level of 75% for the second-stage network into the second-stage network to obtain the second-stage network's output. Based on the difference in noise level between the second-stage network's output and the training target signal data with a relative noise level of 50%, we can determine the loss value of the second-stage network. Subsequently, we can perform separate and joint gradient descent based on the loss value of the second-stage network.

[0137] In one embodiment, before inputting the training input signal data corresponding to the current-level network into the current-level network, the training method of the signal denoising model provided by the present application further includes:

[0138] According to the training input signal data and training target signal data corresponding to the previous level network, the noise removed by the previous level network of the current level network is obtained; the noise removed by the previous level network of the current level network is subjected to signal extraction; and the extracted signal is added to the training input signal data corresponding to the current level network.

[0139] For example, if the second-level noise reduction training is currently in progress and the noise reduction calibration range is divided by the absolute noise level, if the noise reduction calibration range of the second-level network is 300 to 200, then the training input signal data corresponding to the second-level network with an absolute noise level of 300 can be obtained, and the training target signal data corresponding to the second-level network with an absolute noise level of 200 can be obtained. Accordingly, the network before the second-level network is the first-level network.

[0140] The training input signal data corresponding to the first-stage network and the training target signal data can be subtracted to obtain difference signal data. The difference signal data is not much different from the noise actually removed by the first-stage network. Therefore, the difference signal data can be used as the noise removed by the first-stage network. Then, the signal compensation algorithm can be used to extract the signal of effective information from the noise removed by the first-stage network. The signal extraction method can be:

[0141] Method 1: Multiply the noise removed by the first-stage network by the set coefficient to obtain the signal that is valid information in the noise removed by the first-stage network;

[0142] Method 2: Perform low-frequency domain extraction on the noise removed by the first-stage network to obtain the signal that is valid information in the noise removed by the first-stage network;

[0143] Method three: The noise removed by the first-stage network is input into the signal extraction network through deep learning to obtain the signal containing valid information from the noise removed by the first-stage network.

[0144] A set proportion of the signal removed from the noise by the first-stage network is first added to the training input signal data corresponding to the second-stage network, which has an absolute noise level of 300. This training input signal data is then input into the second-stage network to obtain the output of the second-stage network. The loss value of the second-stage network can be obtained based on the difference in noise level between the output of the second-stage network and the training target signal data corresponding to the second-stage network, which has an absolute noise level of 200. Subsequently, separate gradient descent and joint gradient descent can be performed based on the loss value of the second-stage network.

[0145] In this embodiment, between the training of two adjacent levels of networks, the signal in the noise removed by the previous level network is added to the training input signal data corresponding to the current level network, thereby ensuring that excessive signal loss does not occur during the step-by-step noise reduction training process, thereby achieving the purpose of noise reduction under better signal-to-noise ratio conditions.

[0146] In one embodiment, the multi-level networks are networks with the same network structure, or the multi-level networks are networks with different network structures.

[0147] Multi-level networks can have the same network structure or use different network structures based on requirements.

[0148] In one embodiment, when the multi-stage network is a network with different network structures, the structural complexity of the previous stage network in two adjacent stages is higher than the structural complexity of the subsequent stage network, and / or the parameter amount of the previous stage network in two adjacent stages is greater than the parameter amount of the subsequent stage network.

[0149] The following is an example of two adjacent networks, the first-level network and the second-level network.

[0150] The first-stage network is closer to the model input than the second-stage network. The first-stage network is exposed to greater noise, placing a higher requirement on its signal-to-noise ratio maintenance. Therefore, the first-stage network can have a higher structural complexity and / or a larger number of parameters. In contrast, the second-stage network is exposed to less noise, placing a lower requirement on its signal-to-noise ratio maintenance and a higher requirement on its signal restoration capability. Therefore, the second-stage network can have a lower structural complexity and / or a smaller number of parameters.

[0151] In one embodiment, the signal data includes any one of CT raw data and CT images.

[0152] CT stands for Computed Tomography. In signals like CT images, the structure within them is valuable information. Therefore, the signal extracted from the noise removed by the previous network can include this structural information.

[0153] In one embodiment, when the signal data includes CT raw data, and the CT raw data is CT multi-level data, the training input CT multi-level data is merged and input into the network or is input into the network independently;

[0154] In the case where the signal data includes a CT image, and the CT image is a CT multi-energy level image, the training input CT multi-energy level image is merged and input into the network or is independently input into the network.

[0155] In the case where the signal data includes CT raw data, and the CT raw data is CT multi-level data, the training input signal data can be referred to as training input CT multi-level data. The training input CT multi-level data can be merged and input into the network. The merging methods include but are not limited to: (1) merging the training input CT multi-level data from the channel direction, (2) merging the training input CT multi-level data from an independent dimension, such as merging in the z direction. The training input CT multi-level data can also be independently input into the network.

[0156] The training input CT multi-level data is merged into the network or input into the network independently, and the loss value of the input CT multi-level data is subsequently trained together.

[0157] In the case where the signal data includes a CT image, and the CT image is a CT multi-level image, the training input signal data can be referred to as a training input CT multi-level image. The training input CT multi-level image can be merged and input into the network. The merging methods include but are not limited to: (1) merging the training input CT multi-level image from the channel direction, (2) merging the training input CT multi-level image from an independent dimension, such as merging in the z direction. The training input CT multi-level image can also be independently input into the network.

[0158] The training input CT multi-level images are merged into the input network or independently input into the network, and the loss values of the input CT multi-level images are subsequently trained together.

[0159] To better understand the training method of the above-mentioned signal noise reduction model, an application example is described in detail below. In this application example, the signal noise reduction model includes a four-level network, and the noise reduction calibration range is divided by the absolute noise size. Assume that the noise reduction calibration range corresponding to the first-level network is 400 to 300, the noise reduction calibration range corresponding to the second-level network is 300 to 200, the noise reduction calibration range corresponding to the third-level network is 200 to 100, and the noise reduction calibration range corresponding to the fourth-level network is 100 to 0.

[0160] Based on the noise reduction calibration range corresponding to each level of the network, the training signal data corresponding to each level of the network is obtained. Each training signal data includes training input signal data and training target signal data. The noise level carried by the training input signal data is the upper limit of the noise reduction calibration range corresponding to the corresponding level of the network, and the noise level carried by the training target signal data is the lower limit of the noise reduction calibration range corresponding to the corresponding level of the network.

[0161] The training input signal data corresponding to the first-level network is recorded as signal-input(400), and the training target signal data corresponding to the first-level network is recorded as signal-target(300). The training input signal data corresponding to the second-level network is recorded as signal-input(300), and the training target signal data corresponding to the second-level network is recorded as signal-target(200). The training input signal data corresponding to the third-level network is recorded as signal-input(200), and the training target signal data corresponding to the third-level network is recorded as signal-target(100). The training input signal data corresponding to the fourth-level network is recorded as signal-input(100), and the training target signal data corresponding to the fourth-level network is recorded as signal-target(0).

[0162] During the first-level noise reduction training, as shown in FIG2(b), the training input signal data signal-input (400) corresponding to the first-level network can be input into the first-level network. The difference in noise level between the output of the first-level network and the training target signal data signal-target (300) corresponding to the first-level network is calculated using the loss function to obtain the loss value loss-1 of the first-level network. Based on the loss value loss-1 of the first-level network, backpropagation and gradient descent are performed on the first-level network alone to complete the first-level noise reduction training.

[0163] After completing the first-level noise reduction training, the second-level noise reduction training is performed. As shown in Figure 2 (c), the signal in the noise removed by the first-level network can be added to the training input signal data signal-input (300) corresponding to the second-level network, and then the training input signal data signal-input (300) is input into the second-level network; through the loss function, the difference in noise size between the output result of the second-level network and the training target signal data signal-target (200) corresponding to the second-level network is calculated to obtain the loss value loss-2 of the second-level network. Based on the loss value loss-2 of the second-level network, the second-level network is back-propagated and gradient descent is performed separately; after back-propagation and gradient descent, the first-level network and the second-level network are jointly back-propagated and gradient descent based on the loss value loss-2 of the second-level network to complete the second-level noise reduction training.

[0164] After completing the second-level noise reduction training, the third-level noise reduction training is performed. As shown in Figure 2 (d), the signal in the noise removed by the second-level network can be added to the training input signal data signal-input (200) corresponding to the third-level network, and then the training input signal data signal-input (200) is input into the third-level network; through the loss function, the difference in noise size between the output result of the third-level network and the training target signal data signal-target (100) corresponding to the third-level network is calculated to obtain the loss value loss-3 of the third-level network. Based on the loss value loss-3 of the third-level network, the third-level network is back-propagated and gradient descent is performed separately; after back-propagation and gradient descent, the first-level network, the second-level network, and the third-level network are jointly back-propagated and gradient descent based on the loss value loss-3 of the third-level network to complete the third-level noise reduction training.

[0165] After completing the third-level noise reduction training, the fourth-level noise reduction training is performed. As shown in Figure 2 (e), the signal in the noise removed by the third-level network can be added to the training input signal data signal-input (100) corresponding to the fourth-level network, and then the training input signal data signal-input (100) is input into the fourth-level network; through the loss function, the difference in noise size between the output result of the fourth-level network and the training target signal data signal-target (0) corresponding to the fourth-level network is calculated to obtain the loss value loss-4 of the fourth-level network. Based on the loss value loss-4 of the fourth-level network, the fourth-level network is back-propagated and gradient descent is performed separately; after back-propagation and gradient descent, based on the loss value loss-4 of the fourth-level network, the first-level network, the second-level network, the third-level network, and the fourth-level network are jointly back-propagated and gradient descent to complete the fourth-level noise reduction training.

[0166] After the fourth level of noise reduction training is completed, one iteration is considered complete. Two or more iterations can be performed based on the actual situation. After the iteration is completed, the signal noise reduction model is obtained.

[0167] When the signal data is CT multi-level data, if the number of energy levels is two, the architecture diagram of the first-level noise reduction training is shown in Figure 3(a), and the architecture diagram of the second-level noise reduction training is shown in Figure 3(b). The architecture diagrams of the third-level noise reduction training and the fourth-level noise reduction training are not shown.

[0168] The signal denoising method provided in this application belongs to the inference stage. The signal denoising method can be executed by a computer device, and the signal denoising method may include Figure 4 Steps shown.

[0169] Step S401: Acquire a signal to be denoised.

[0170] If the signal in the training phase is CT raw data, the CT raw data to be denoised can be obtained in the inference phase. If the signal in the training phase is CT images, the CT images to be denoised can be obtained in the inference phase.

[0171] Step S402: input the signal to be denoised into a signal denoising model to obtain a denoised signal.

[0172] The signal denoising model is obtained by training according to the signal denoising model training method introduced in the above embodiment.

[0173] After obtaining the signal to be denoised, the signal to be denoised can be input into a signal denoising model, and a denoised signal can be obtained according to an output result of the signal denoising model.

[0174] The signal denoising model adopted in this embodiment is obtained through a step-by-step denoising training method; during the training process, not only a single-stage network is subjected to a separate gradient descent, but also a multi-stage network is subjected to a joint gradient descent, which can improve the denoising ability of the signal denoising model; and, in the training signal data corresponding to the current-stage network, the signal in the noise removed by the previous-stage network of the current-stage network is added to compensate and restore the signal, thereby avoiding the situation where the signal is weakened after denoising at each stage of the network during the training process, so that the signal denoising model can perform denoising while ensuring signal accuracy, so that the denoised signal obtained contains a higher signal accuracy and retains more signals.

[0175] In one embodiment, the signal to be denoised is input into a signal denoising model to obtain a denoised signal, including:

[0176] Determine the noise level of the signal to be denoised; determine the target input network in the signal denoising model according to the noise level of the signal to be denoised; the noise level of the signal to be denoised is within the noise reduction calibration range of the target input network; input the signal to be denoised into the target input network, so that the forward multi-stage network starting from the target input network in the signal denoising model performs step-by-step denoising on the signal to be denoised, thereby obtaining a denoised signal.

[0177] The forward direction (also called positive direction) is the direction from the model input end to the model output end, and correspondingly, the backward direction (also called reverse direction) is the direction from the model output end to the model input end.

[0178] The noise reduction calibration range is divided by the absolute noise size. The signal noise reduction model includes a four-level network. For example, the noise reduction calibration range corresponding to the first-level network is 400 to 300, the noise reduction calibration range corresponding to the second-level network is 300 to 200, the noise reduction calibration range corresponding to the third-level network is 200 to 100, and the noise reduction calibration range corresponding to the fourth-level network is 100 to 0.

[0179] If the noise level of the signal to be denoised is 250, 250 is within the noise reduction calibration range of 300 to 200, so the second-level network can be used as the target input network. The forward multi-level network starting with the second-level network includes the second-level network, the third-level network and the fourth-level network. Figure 5 As shown in the figure, after the signal to be denoised is input into the second-stage network, it is subjected to step-by-step denoising by the second, third, and fourth-stage networks. The denoised signal is obtained based on the output of the fourth-stage network. The noise level of the signal to be denoised can be estimated using a noise extraction algorithm.

[0180] The noise reduction calibration range is divided by the relative noise size. The signal noise reduction model includes a four-level network. For example, the noise reduction calibration range corresponding to the first-level network is 100% to 75%, the noise reduction calibration range corresponding to the second-level network is 75% to 50%, the noise reduction calibration range corresponding to the third-level network is 50% to 25%, and the noise reduction calibration range corresponding to the fourth-level network is 25% to 0.

[0181] If the calibrated noise value is 500 and the noise level of the signal to be denoised is 300, the relative noise level of the signal to be denoised relative to the calibrated noise value is 60%. 60% is within the noise reduction calibration range of 75% to 50%, so the second-level network can be used as the target input network. The forward multi-stage network starting with the second-level network includes the second-level network, the third-level network, and the fourth-level network. Figure 5 As shown in the figure, after the signal to be denoised is input into the second-stage network, it is subjected to step-by-step denoising by the second, third, and fourth-stage networks. The denoised signal is obtained based on the output of the fourth-stage network. The noise level of the signal to be denoised can be estimated using a noise extraction algorithm.

[0182] In this embodiment, during inference, it is possible to dynamically determine which level of network to input based on the noise level of the signal to be denoised, thereby obtaining results more quickly and improving the denoising efficiency.

[0183] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0184] Based on the same inventive concept, an embodiment of the present application also provides a signal denoising model training device for implementing the signal denoising model training method involved above. The implementation solution provided by this device is similar to the implementation solution described in the above method. Therefore, the specific limitations of the one or more signal denoising model training device embodiments provided below can be found in the above-mentioned limitations of the signal denoising model training method, and will not be repeated here.

[0185] This application provides a training device for a signal denoising model, comprising:

[0186] A model acquisition module is used to perform noise reduction training on the multi-level network step by step based on the training signal data corresponding to the multi-level network to obtain a signal noise reduction model;

[0187] The training module is used to obtain the loss value of the current-level network according to the training signal data corresponding to the current-level network if the current level trained by the step-by-step noise reduction is not the first level; after performing gradient descent on the current-level network based on the loss value, the current multi-level network that has been gradient descended separately is jointly gradient descended based on the loss value.

[0188] In one embodiment, the training signal data corresponding to the current level network is added with the signal in the noise removed by the previous level network of the current level network.

[0189] In one embodiment, the training module is further configured to:

[0190] Based on the loss value, the current-level network is reversely calculated layer by layer to obtain a first-type gradient value; based on the first-type gradient value, the current-level network is gradient descended.

[0191] In one embodiment, the training module is further configured to:

[0192] Based on the loss value, reverse calculation is performed on the current multi-level individual gradient descent network network to obtain a second type of gradient value; based on the second type of gradient value, gradient descent is performed on the current multi-level individual gradient descent network.

[0193] In one embodiment, the signal denoising model training device further includes a sample acquisition module for:

[0194] According to the noise reduction calibration range corresponding to each level of the network, the training signal data corresponding to each level of the network is obtained; each training signal data includes training input signal data and training target signal data, the noise size carried by the training input signal data is the upper limit value of the noise reduction calibration range corresponding to the corresponding level of the network, and the noise size carried by the training target signal data is the lower limit value of the noise reduction calibration range corresponding to the corresponding level of the network.

[0195] The training module is also used to:

[0196] The training input signal data corresponding to the current level network is input into the current level network; and the loss value of the current level network is obtained according to the difference in noise size between the output result of the current level network and the training target signal data corresponding to the current level network.

[0197] In one embodiment, the training module is further configured to:

[0198] According to the training input signal data and training target signal data corresponding to the previous level network, the noise removed by the previous level network of the current level network is obtained; the noise removed by the previous level network of the current level network is subjected to signal extraction; and the extracted signal is added to the training input signal data corresponding to the current level network.

[0199] In one embodiment, the multi-level networks are networks with the same network structure, or the multi-level networks are networks with different network structures.

[0200] In one embodiment, when the multi-stage network is a network with different network structures, the structural complexity of the previous stage network in two adjacent stages is higher than the structural complexity of the subsequent stage network, and / or the parameter amount of the previous stage network in two adjacent stages is greater than the parameter amount of the subsequent stage network.

[0201] In one embodiment, the signal data includes any one of CT raw data and CT images.

[0202] In one embodiment, when the signal data includes CT raw data, when the CT raw data is CT multi-energy level data, the training input CT multi-energy level data is merged into the input network or is independently input into the network; when the signal data includes CT images, when the CT images are CT multi-energy level images, the training input CT multi-energy level images are merged into the input network or are independently input into the network.

[0203] Each module in the aforementioned signal denoising model training device can be implemented in whole or in part via software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0204] Based on the same inventive concept, embodiments of the present application also provide a signal noise reduction device for implementing the aforementioned signal noise reduction method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more signal noise reduction device embodiments provided below can be found in the aforementioned limitations of the signal noise reduction method and will not be further elaborated here.

[0205] The present application provides a signal noise reduction device, comprising:

[0206] A signal acquisition module, used to acquire the signal to be denoised;

[0207] The denoising module is used to input the signal to be denoised into a signal denoising model to obtain a denoised signal; the signal denoising model is trained according to the signal denoising model training method introduced in the above embodiment.

[0208] In one embodiment, the noise reduction module is further configured to:

[0209] Determine the noise level of the signal to be denoised; determine a target input network in a signal denoising model based on the noise level of the signal to be denoised; the noise level of the signal to be denoised is within a noise reduction calibration range of the target input network; input the signal to be denoised into the target input network, so that a forward multi-stage network starting from the target input network in the signal denoising model performs step-by-step denoising on the signal to be denoised to obtain a denoised signal.

[0210] Each module in the above-mentioned signal noise reduction device can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a memory in a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0211] In an exemplary embodiment, a computer device is provided, the internal structure of which can be shown as follows: Figure 6 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store data involved in the above method. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a method introduced in the above embodiment.

[0212] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0213] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0214] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0215] In one embodiment, a computer program product is provided, on which a computer program is stored. The computer program is used by a processor to execute the steps in the above-mentioned various method embodiments.

[0216] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0217] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), quantum computing-based data processing logic devices, artificial intelligence (AI) processors, and the like.

[0218] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0219] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A method for training a signal denoising model, characterized in that: The method comprises: Based on the training signal data corresponding to the multi-level network, the multi-level network is trained for noise reduction step by step to obtain a signal noise reduction model; Wherein, if the current level trained by the step-by-step noise reduction is not the first level, the loss value of the current level network is obtained according to the training signal data corresponding to the current level network; After performing gradient descent on the current level network individually based on the loss value, gradient descent is jointly performed on the current multi-level networks that have already undergone individual gradient descent based on the loss value.

2. The method according to claim 1, characterized in that The training signal data corresponding to the current level network is added with the signal in the noise removed by the previous level network of the current level network.

3. The method according to claim 1, characterized in that Performing gradient descent on the current level network based on the loss value, including: Based on the loss value, reversely calculate the current level network layer by layer to obtain a first type of gradient value; Based on the first type of gradient value, gradient descent is performed on the current level network.

4. The method according to claim 1, wherein Based on the loss value, the current multi-level gradient descent network is jointly gradient-descented, including: Based on the loss value, reverse calculation is performed on the current multi-level individual gradient descent network one by one to obtain a second type of gradient value; Based on the second type of gradient value, gradient descent is performed on the current multi-level individual gradient descent network.

5. The method according to claim 1, wherein The method further comprises: According to the noise reduction calibration range corresponding to each level of the network, the training signal data corresponding to each level of the network is obtained; each training signal data includes training input signal data and training target signal data, the noise level carried by the training input signal data is the upper limit value of the noise reduction calibration range corresponding to the corresponding level of the network, and the noise level carried by the training target signal data is the lower limit value of the noise reduction calibration range corresponding to the corresponding level of the network; Obtaining a loss value of the current-level network based on the training signal data corresponding to the current-level network includes: Inputting the training input signal data corresponding to the current level network into the current level network; The loss value of the current-level network is obtained according to the difference in noise level between the output result of the current-level network and the training target signal data corresponding to the current-level network.

6. The method according to claim 5, characterized in that Before inputting the training input signal data corresponding to the current-level network into the current-level network, the method further includes: Obtaining noise removed by a previous-stage network of the current-stage network based on training input signal data and training target signal data corresponding to a previous-stage network; Performing signal extraction on the noise removed by the previous level network of the current level network; The extracted signal is added to the training input signal data corresponding to the current level network.

7. The method according to claim 1, characterized in that The signal data includes either CT raw data or CT images.

8. A signal noise reduction method, characterized in that: The method comprises: Obtaining a signal to be denoised; The signal to be denoised is input into a signal denoising model to obtain a denoised signal; the signal denoising model is trained according to the signal denoising model training method according to any one of claims 1 to 7.

9. The method according to claim 8, characterized in that Inputting the signal to be denoised into a signal denoising model to obtain a denoised signal includes: Determining the noise level of the signal to be denoised; Determining a target input network in a signal noise reduction model according to the noise level of the signal to be denoised; the noise level of the signal to be denoised is within a noise reduction calibration range of the target input network; The signal to be denoised is input into the target input network, so that the forward multi-stage network starting from the target input network in the signal denoising model performs step-by-step denoising on the signal to be denoised to obtain a denoised signal.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 9 is implemented.

Images