Polarization code image signal decoding method for indirect noise estimation by using CNN and YOLOv4

Abstract

The invention discloses a polarization code image signal decoding method for indirect noise estimation by using CNN and YOLOv4, which comprises the following steps: pre-judging a received noisy signal to generate a pre-judging binary sequence, and subtracting the pre-judging binary sequence from the noisy signal to obtain a preliminary noise estimation signal; sending the preliminary noise estimation signal to an improved CNN for noise feature extraction to obtain a noise feature signal, and subtracting the noise feature signal from a noisy signal to obtain a noise suppression signal; and the noise suppression signal is sent to the improved YOLOv4 for noise detection to obtain a noise detection signal, the noise suppression signal and the noise detection signal are spliced to obtain a fusion signal, and the fusion signal is decoded by using an SCAN decoding algorithm to obtain a restored signal. According to the method, the bit error rate under complex interference is reduced by 40%-50% while the decoding efficiency is kept, the adaptive capacity of a decoding algorithm to noise is effectively improved, and the method is remarkably superior to a method for directly optimizing polar code decoding through deep learning.

Description

Technical Field

[0001] The present invention relates to the technical field of image decoding, and in particular to a polar code image signal decoding method using CNN and YOLOv4 for indirect noise estimation. Background Art

[0002] Polar codes, characterized by simple coding and high reliability, have been adopted as the control channel coding scheme for 5G wireless communications. Traditional signal decoding methods based on polar code encoding schemes primarily employ the successive cancellation (SC) decoding algorithm, the serial cancellation list (SCL) decoding algorithm, and the probability-based iterative belief propagation (BP) decoding algorithm. The SC decoding algorithm primarily relies on hard decisions and cannot effectively utilize soft information, resulting in insufficient error correction capabilities for burst errors and mixed noise. Furthermore, its serial decoding structure limits decoding speed, making it difficult to meet real-time requirements. The computational complexity of the SC decoding algorithm increases exponentially with the size of the list. Furthermore, it is still based on hard decisions, with limited dynamic adjustment capabilities for soft information. This performance improvement is particularly limited in low signal-to-noise ratio (SNR) or non-stationary noise environments. The probability-based iterative belief propagation decoding algorithm has slow convergence and is prone to falling into local optima under poor channel conditions. It also requires multiple iterations (typically 10-20), consuming large amounts of computational resources and making it unsuitable for real-time transmission scenarios.

[0003] Compared to signal decoding methods based on polar code encoding, the Soft Cancellation-Assisted N-Path Scanning (SCAN) decoding algorithm significantly improves decoding performance by introducing a soft information iterative update mechanism and a multi-path scanning strategy. Soft information iteration allows the algorithm to dynamically adjust the log-likelihood ratio (LLR), preserving the uncertainty of channel noise. Multi-path scanning maintains multiple candidate decoding paths, reducing the risk of overall failure due to misjudgment of a single path. However, existing SCAN decoding algorithms primarily focus on communication signal processing, and their application in the image domain is immature, particularly in complex channel interference scenarios, which severely hinders the upper limit of SCAN decoding performance.

[0004] Given that the decoding algorithms used in current signal decoding methods based on polar code encoding suffer from issues such as large decoding delays and high iteration complexity, deep learning technology has been introduced to address these bottlenecks. However, existing research has primarily focused on using deep learning to directly optimize the internal decoding process of the decoding algorithm. For example, this involves dynamically correcting the log-likelihood ratio (LLR) during decoding by building neural network models, or optimizing the path search strategy during decoding by modeling the state transition relationships in sequence decoding using recurrent neural networks (RNNs). However, this method of using deep learning to directly optimize the internal decoding process of the decoding algorithm has its shortcomings: first, the fusion of the neural network model and the decoding algorithm remains at a shallow level, failing to achieve deep coordination between the channel estimation and decoding processes, resulting in the accumulation and amplification of noise estimation errors during decoding iterations; second, for complex interference scenarios (such as mixed environments of impulse noise and Gaussian noise), existing neural network models rely on feature extraction layers with fixed structures, which are insufficiently adaptable to multimodal noise, and when processing high-resolution image domain signals, there is a problem of synchronous growth in computational complexity and decoding delay; third, when the neural network model is directly applied to noise classification, due to the high redundancy of network parameters, it is difficult to meet the computing power limitations of low-power devices in real-time transmission scenarios, resulting in limited engineering practicality of the decoding system. Summary of the Invention

[0005] The present invention aims to solve the problems existing in the existing method of directly optimizing the internal decoding process of the decoding algorithm using deep learning, and provides a polar code image signal decoding method using CNN and YOLOv4 for indirect noise estimation.

[0006] To solve the above problems, the present invention is achieved through the following technical solutions:

[0007] A polar code image signal decoding method using CNN and YOLOv4 for indirect noise estimation includes the following steps:

[0008] Step 1: pre-judge the received noisy signal y to generate a pre-judgement binary sequence s1, and subtract the pre-judgement binary sequence s1 from the noisy signal y to obtain a preliminary noise estimation signal n1;

[0009] Step 2: The preliminary noise estimation signal n1 is fed into the improved CNN for noise feature extraction to obtain the noise feature signal n2, and the noise feature signal n2 is subtracted from the noisy signal y to obtain the noise suppression signal y1;

[0010] Step 3: Send the noise suppression signal y1 to the improved YOLOv4 for noise detection to obtain the noise detection signal n3, and then concatenate the noise suppression signal y1 with the noise detection signal n3 to obtain the fusion signal y2;

[0011] Step 4: Use the SCAN decoding algorithm to decode the fused signal y2 to obtain the restored signal u1.

[0012] The above-mentioned improved CNN consists of an input layer, 4 convolutional layers, 4 activation function layers and an output layer; the input of the input layer serves as the input of the improved CNN; the output of the input layer is connected to the input of the first convolutional layer, the output of the first convolutional layer is connected to the input of the first activation function layer, the output of the first activation function layer is connected to the input of the second convolutional layer, the output of the second convolutional layer is connected to the input of the second activation function layer, the output of the second activation function layer is connected to the input of the third convolutional layer, the output of the third convolutional layer is connected to the input of the third activation function layer, the output of the third activation function layer is connected to the input of the fourth convolutional layer, the output of the fourth convolutional layer is connected to the input of the fourth activation function layer, and the output of the fourth activation function layer is connected to the input of the output layer; the output of the output layer serves as the output of the improved CNN.

[0013] The improved YOLOv4 consists of an input layer, a two-dimensional convolutional layer, a batch normalization layer, an activation function layer, 4 residual modules, a spatial pyramid pooling module, 2 convolutional layers, 2 three-level convolutional modules, 4 five-level convolutional modules, 2 upsampling layers, 2 downsampling layers, 4 splicing layers, 3 detection heads and an output layer; the input of the input layer is used as the input of the improved YOLOv4; the output of the input layer is connected to the input of the two-dimensional convolutional layer, the output of the two-dimensional convolutional layer is connected to the input of the batch normalization layer, the output of the batch normalization layer is connected to the input of the activation function layer, the output of the activation function layer is connected to the input of the first residual module; the output of the first residual module is connected The output of the second residual module is connected to the input of the third residual module, the output of the third residual module is connected to the input of the fourth residual module, and the output of the fourth residual module is connected to the input of the first three-level convolution module; the output of the first three-level convolution module is connected to the input of the spatial pyramid pooling module, the output of the spatial pyramid pooling module is connected to the input of the second three-level convolution module, the output of the second three-level convolution module is connected to the input of the first convolution layer, the output of the first convolution layer is connected to the input of the first upsampling layer, and the output of the first upsampling layer and the output of the third residual module are simultaneously connected to the first splicing layer. Input, the output of the first splicing layer is connected to the input of the first five-level convolution module, the output of the first five-level convolution module is connected to the input of the second convolution layer, the output of the second convolution layer is connected to the input of the second upsampling layer, the output of the second upsampling layer and the output of the second residual module are simultaneously connected to the input of the second splicing layer, the output of the second splicing layer is connected to the input of the second five-level convolution module, the output of the second five-level convolution module is connected to the input of the first downsampling, the output of the first downsampling and the output of the first five-level convolution module are simultaneously connected to the input of the third splicing layer, and the output of the third splicing layer is connected to the input of the third five-level convolution module. The output of the third five-level convolution module is connected to the input of the second downsampling, the output of the second downsampling and the output of the second three-level convolution module are connected to the input of the fourth splicing layer at the same time, and the output of the fourth splicing layer is connected to the input of the fourth five-level convolution module; the output of the second five-level convolution module is connected to the input of the first detection head, the output of the third five-level convolution module is connected to the input of the second detection head, and the output of the fourth five-level convolution module is connected to the input of the third detection head; the output of the first detection head, the output of the second detection head and the output of the third detection head are connected to the input of the output layer at the same time; the output of the output layer is used as the output of the improved YOLOv4.

[0014] In the above-mentioned improved YOLOv4, the first residual module includes 1 residual module, the second residual module and the third residual module each include 8 repeated residual modules, and the fourth residual module includes 4 repeated residual modules.

[0015] The present invention comprehensively uses an improved CNN (convolutional neural network) and an improved YOLOv4 (fourth version single-stage object detection network) to extract and estimate noise features, and realizes efficient decoding through the SCAN (soft cancellation assisted N-path scanning) decoding algorithm. It does not use deep learning to directly optimize the internal decoding process of the decoding algorithm, but uses deep learning to estimate the channel noise to achieve the purpose of indirectly optimizing the decoding algorithm.

[0016] Compared with the prior art, the present invention has the following characteristics:

[0017] 1. Multimodal feature complementarity: CNN's signal amplitude optimization and YOLOv4's noise category localization form a hierarchical fusion of "low-level features and high-level semantics," enabling SCAN decoding to simultaneously utilize signal waveform details (such as the continuous fluctuations of Gaussian noise) and noise distribution priors (such as the pulse positions of salt and pepper noise). This reduces the bit error rate by 35% to 45% compared to directly decoding using the SCAN algorithm, and by 20% to 30% compared to directly optimizing the decoding process through deep learning.

[0018] 2. Reduced decoding complexity: By extracting noise features through a lightweight CNN and parallel feature processing, its computational complexity is reduced by 50% compared to traditional image denoising networks. In addition, the noise features output by the CNN are combined with the noise type and location output by YOLOv4 to achieve accurate noise estimation, allowing SCAN decoding to adaptively adjust the number of paths and iterations, maintaining equivalent error correction performance in complex noise scenarios. The computational complexity is reduced by 40%, and the single-frame signal processing delay is ≤80μs (code length N=1024), meeting the 1ms transmission latency requirement for 5G communications.

[0019] 3. Enhanced robustness: YOLOv4 detects noise type changes in real time (such as a sudden change from Gaussian noise to composite noise). SCAN decoding can complete parameter switching within two symbol periods, reducing the bit error rate fluctuation by 70% compared to traditional fixed-parameter algorithms.

[0020] 4. End-to-end collaborative optimization: The fused signal output by CNN-YOLOv4 is used as an intermediate training variable to support end-to-end optimization of CNN-YOLOv4-SCAN. This shifts the system's bit error rate curve by 1.5 dB to the left in low signal-to-noise ratio (SNR ≤ 5dB) scenarios compared to traditional deep learning cascade solutions, further improving overall system performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The schematic diagram of the image transmission process based on polar code encoding.

[0022] Figure 2 Schematic diagram of the improved CNN.

[0023] Figure 3 Schematic diagram of the improved YOLOv4. DETAILED DESCRIPTION

[0024] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific examples and the accompanying drawings.

[0025] The image transmission process based on polar code encoding is as follows Figure 1 As shown, it mainly includes the decoding process of the transmitting end and the decoding process of the receiving end.

[0026] At the image transmitter: First, a CCD camera and infrared sensor are used for multimodal image acquisition to obtain raw image information containing the target object. This information is converted into an electrical signal u that carries information related to the target image. The waveform is a continuous voltage signal, whose amplitude corresponds to the pixel brightness (0-5V, corresponding to the ADC full-scale voltage), and the structure is the line scan signal of a two-dimensional image (each row of pixels is output in a time series). Then, polar code encoding is used with a code length N = 1024, information bits K = 512, and check bits NK = 512. The analog signal u is converted into a digital signal by a 16-bit ADC (AD9250, sampling rate 20MSPS). Polar code encoding is then used to generate the polar code codeword x, which converts the continuous analog signal into a discrete binary sequence. Check bits are added to prevent transmission errors. After conversion, the codeword is suitable for subsequent transmission and processing. Next, binary phase-shift keying (BPSK) maps each bit of the polar codeword x into a phase-modulated signal. Bit 0 is mapped to a carrier phase of 0°, with a waveform of A cosine (2πft), and bit 1 is mapped to a carrier phase of 180°, with a waveform of -A cosine (2πft). A is assumed to be 5V (ADC full-scale voltage), and the carrier frequency is set according to the channel bandwidth. This generates the analog signal s, a sequence of phase-hopping sinusoidal waves with a symbol period of T = 1 / f = 100ns. This structure is a continuous phase-modulated waveform carrying the polar code-encoded information. Finally, the modulated signal s enters the channel.

[0027] In the channel, multiple noises mix and interfere with the signal s, causing random fluctuations in the waveform amplitude. For example, in some places, the original signal changes from +A to +A+n(t), or the signal suddenly jumps to +5V or -5V, resulting in single or continuous bit errors (such as multiple symbols being overwritten by pulses). In some places, the waveform is subjected to both continuous perturbations of Gaussian noise and sudden impacts of random pulses, resulting in a significant increase in the bit error rate. The error distribution includes both random single-bit errors and sudden clustered errors. After the signal is interfered with, it is converted into a noisy signal y.

[0028] At the receiving end of the image, the method proposed in the present invention is used to decode the noisy signal y to restore the signal sent by the transmitting end, thereby obtaining the restored signal u1. Specifically, the polar code image signal decoding method proposed in the present invention using CNN and YOLOv4 for indirect noise estimation includes the following steps:

[0029] Step 1: First, pre-judge the received noisy signal y to generate a pre-judgement binary sequence s1. Then, subtract the pre-judgement binary sequence s1 from the noisy signal y to obtain a preliminary noise estimation signal n1 = y-s1.

[0030] Pre-determination is achieved using a sampling decision maker. This is a key component that recovers or regenerates baseband signals in the presence of imperfect transmission characteristics and noise, using a sampling and decision process controlled by bit timing pulses. If the signal amplitude y(t) within the current symbol period is greater than 0, the decision is made as bit 0, corresponding to a theoretical transmission waveform of +5V (full-scale voltage). If the signal amplitude y(t) within the current symbol period is less than or equal to 0, the decision is made as bit 1, corresponding to a theoretical transmission waveform of -5V.

[0031] The pre-decision binary sequence s1=[b1,T2,…,b N ]; where b i ∈{0,1}, corresponding to the i-th symbol period of the noisy signal y; N = 1024 is the polar code length. Assuming that the pre-decision binary sequence s1 output is a theoretically noise-free BPSK signal, subtract it from the actual received signal y to obtain the preliminary noise estimate signal n1.

[0032] Step 2: First, the preliminary noise estimation signal n1 is fed into the improved CNN for noise feature extraction to obtain the noise feature signal n2. Then, the noise feature signal n2 is subtracted from the noisy signal y to obtain the noise suppression signal y1 = y - n2.

[0033] See also Figure 2 The improved CNN, i.e., lightweight CNN, consists of an input layer, four convolutional layers, four activation function layers, and an output layer. The input of the input layer serves as the input of the improved CNN; the output of the input layer is connected to the input of the first convolutional layer, the output of the first convolutional layer is connected to the input of the first activation function layer, the output of the first activation function layer is connected to the input of the second convolutional layer, the output of the second convolutional layer is connected to the input of the second activation function layer, the output of the second activation function layer is connected to the input of the third convolutional layer, the output of the third convolutional layer is connected to the input of the third activation function layer, the output of the third activation function layer is connected to the input of the fourth convolutional layer, the output of the fourth convolutional layer is connected to the input of the fourth activation function layer, and the output of the fourth activation function layer is connected to the input of the output layer; the output of the output layer serves as the output of the improved CNN.

[0034] Unlike the network structure used for low-level tasks in image denoising and super-resolution, the present invention converts the preliminary noise estimation signal n1 into a one-dimensional vector in the input layer to avoid redundant calculations of the two-dimensional image network. While maintaining the noise feature extraction capability, the computational complexity is reduced by 50% compared to the traditional image denoising network. Then, four convolutional layers are used to extract noise features, where the first convolutional layer includes 64 9×1 convolution kernels for extracting wide-range noise features; the second convolutional layer includes 32 3×1 convolution kernels for refining local noise patterns; the third convolutional layer includes 6 3×1 convolution kernels for further compressing features; and the fourth convolutional layer includes 1 15×1 convolution kernel for obtaining the final noise feature signal. The four activation function layers learn the distinguishing features between noise and signal through the ReLU activation function σ(x)=max(0,x) nonlinear transformation. In CNN, the calculation of the feature map follows the following formula:

[0035]

[0036] Where σ is the ReLU activation function, which sets the input value less than 0 to 0 and the value greater than 0 to remain unchanged. The expression is σ(x) = max(0, x); Represents the value of the feature map element at position (i, j) in the lth convolutional layer; is the feature map element at position (i, j) in the l-1th layer; is the weight of the kth convolution kernel in the lth layer; M is the number of feature maps in the previous layer; b (l) is the bias term of the lth layer, which is used to adjust the activation threshold of the neuron; * represents the convolution operation.

[0037] By designing a suitable multi-layer convolutional structure, the lightweight CNN in this paper is capable of multi-scale feature extraction from image noise. Leveraging its efficient computational performance and powerful feature extraction capabilities, the lightweight CNN accurately captures everything from macroscopic noise distribution characteristics to microscopic local noise details. Furthermore, when processing Gaussian noise, the lightweight CNN not only identifies the overall noise distribution range but also accurately extracts subtle characteristic variations of noise points.

[0038] The noise signature signal n2 is then subtracted from the actual received noisy signal y to optimize the noisy signal y, making its waveform closer to the original signal s (i.e., the signal before the noise was added). The noise components are suppressed (for example, the amplitude fluctuation of Gaussian noise is reduced and salt-and-pepper pulses are filtered out). In the frequency domain, the energy of high-frequency noise and low-frequency clutter is significantly reduced, and in the spatial domain, image details are clearer, with reduced blur and noise.

[0039] Step 3: First, the noise suppression signal y1 is fed into the improved YOLOv4 for noise detection to obtain the noise detection signal n3. Then, the noise suppression signal y1 and the noise detection signal n3 are concatenated to obtain the fused signal y2 = {y1, n3}.

[0040] By extracting and optimizing noise features from the noisy signal y using an improved CNN, the noise is suppressed to a certain extent. Directly decoding the signal using SCAN would be very complex, and it does not clearly distinguish between the specific types and locations of noise. Therefore, before decoding, it is fed into the improved YOLOv4 for noise detection. The original YOLOv4, which handles the detection of 80 different targets, requires a network parameter count of 63.98M. This requires learning a large amount of information about the features, morphology, and texture of various targets, which requires the network to possess strong feature representation capabilities and, consequently, a large number of parameters. The present invention utilizes YOLOv4 to detect noise in the target, including Gaussian, salt-and-pepper, and composite noise. Since the model has fewer categories to learn, the learning burden is significantly reduced, and fewer parameters are needed to distinguish between different target categories. Therefore, using such a large number of parameters is somewhat wasteful when handling noise detection tasks. To address this issue, the present invention improves the original YOLOv4 network structure, reducing network parameters without reducing the detection model's mean average performance (MAP).

[0041] See also Figure 3 The improved YOLOv4 consists of an input layer, a CSPDarknet48 structure, a PANet structure, a YOLO Head structure, and an output layer. The input layer's input serves as the input of the improved YOLOv4. The output of the output layer serves as the output of the improved YOLOv4. The CSPDarknet48 structure, the PANet structure, and the YOLO Head structure are located between the input and output layers.

[0042] The noise suppression signal y1 is first converted to a format suitable for YOLOv4 input at the input layer. During YOLOv4's forward propagation, the noise suppression signal y1 passes through the CSPDarknet48 structure, the PANet structure, and the YOLO Head structure. In the CSPDarknet48 structure, the signal undergoes a series of convolution, activation, and residual connection operations to extract features at different levels. The PANet structure fuses high-level semantic features with low-level detail features. The YOLO Head structure predicts the location and category of the noise based on the fused features by matching and calculating with the new prior box. Finally, YOLOv4 outputs the processed result through the noise feature branch as the noise detection signal n3. The noise detection signal n3 is a vector containing information about the noise category and may also contain information such as the noise location and confidence level.

[0043] In the CSPDarknet48 structure, the output of the input layer is connected to the input of the two-dimensional convolutional layer, the output of the two-dimensional convolutional layer is connected to the input of the batch normalization layer, the output of the batch normalization layer is connected to the input of the activation function layer, and the output of the activation function layer is connected to the input of the first residual module; the output of the first residual module is connected to the input of the second residual module, the output of the second residual module is connected to the input of the third residual module, the output of the third residual module is connected to the input of the fourth residual module, and the output of the fourth residual module is connected to the input of the first three-level convolutional module.

[0044] The two-dimensional convolution layer (DarknetConv2D) is a two-dimensional convolution layer in the Darknet framework, which is used to extract image features. The batch normalization layer (BN) is used to normalize the data of each batch, speed up the convergence of the model, and alleviate the problem of gradient disappearance. The activation function layer (Mish) adopts the Mish activation function, which has the characteristics of no upper bound, lower bound, smoothness, etc., which allows the model to learn features better. The residual module (Resblock_body) contains multiple convolution layers and cross-layer connections (residual connections). Through cross-layer connections, the gradient is more easily propagated, which helps to train deep networks and alleviate the problem of gradient disappearance. In the present invention, the first residual module includes 1 residual module; the second residual module and the third residual module each include 8 repeated (series) residual modules; the fourth residual module includes 4 repeated (series) residual modules.

[0045] In order to meet the requirements of noise detection for detailed information, this paper reduces the number of network layers of the original YOLOv4 backbone network CSPDarknet53 and no longer uses the original YOLOv4 network depth, but uses CSPDarknet48 instead. Figure 3 As can be seen, we removed the last set of Resblock_body in the original structure. To ensure effective noise feature extraction while reducing parameters, we adjusted the number of Resblock_body in the penultimate set to 4 and the number of Resblock_body in the second set to 8. Furthermore, the number of output channels in the last layer of the backbone network was reduced from 1024 to 512.

[0046] In the PANet structure, the output of the first three-level convolution module is connected to the input of the spatial pyramid pooling module, the output of the spatial pyramid pooling module is connected to the input of the second three-level convolution module, the output of the second three-level convolution module is connected to the input of the first convolution layer, the output of the first convolution layer is connected to the input of the first upsampling layer, the output of the first upsampling layer and the output of the third residual module are simultaneously connected to the input of the first splicing layer, the output of the first splicing layer is connected to the input of the first five-level convolution module, the output of the first five-level convolution module is connected to the input of the second convolution layer, the output of the second convolution layer is connected to the input of the second upsampling layer, and the second upsampling layer is connected to the input of the first convolution layer. The output of the sampling layer and the output of the second residual module are connected to the input of the second splicing layer at the same time, the output of the second splicing layer is connected to the input of the second five-level convolution module, the output of the second five-level convolution module is connected to the input of the first downsampling, the output of the first downsampling and the output of the first five-level convolution module are connected to the input of the third splicing layer at the same time, the output of the third splicing layer is connected to the input of the third five-level convolution module, the output of the third five-level convolution module is connected to the input of the second downsampling, the output of the second downsampling and the output of the second three-level convolution module are connected to the input of the fourth splicing layer at the same time, and the output of the fourth splicing layer is connected to the input of the fourth five-level convolution module.

[0047] Given the uncertainty of noise size, shape, and location, the Spatial Pyramid Pooling (SPP) module uses three different-scale pooling kernels for pooling, processing feature maps from different perspectives. Large-scale pooling kernels focus on the overall characteristics of the noise, while small-scale pooling kernels focus on local details. The three-level convolution module (Convolution×3) achieves feature refinement and consists of three repeated (cascaded) convolutional layers. The first convolution layer performs preliminary processing on the input feature map, specifically extracting basic noise-related features such as initial texture and edge information. The second convolution layer further explores more complex noise feature patterns, such as the unique characteristic combinations of different noise types in specific environments. The third convolution layer integrates and optimizes the features extracted by the first two convolutions, resulting in an output feature map that more accurately represents noise characteristics. The five-level convolution module (Convolution×5) achieves deeper feature mining and consists of five repeated (cascaded) convolutional layers. The five-level convolution module extracts higher-level and more abstract noise features from the multiple fused and processed feature maps. Since noise exhibits complex characteristics in different environments and conditions, these five convolutions can help the model capture these deep-level feature changes and enhance the semantic information and discrimination of the features, allowing the model to more accurately judge the category and location of noise when faced with complex noise scenes. The concatenation layer effectively integrates noise feature information at different levels and scales by splicing multi-scale noise features, and can simultaneously utilize the advantages of high-level semantic features and low-level detail features. The upsampling layer increases the size of the feature map so that the high-level semantic feature map can match the size of the low-level detail feature map, facilitating fusion and helping the model better capture small-scale noise features and improve small noise detection capabilities. The downsampling layer reduces the resolution of the feature map, reducing the amount of computation, extracting more abstract noise features, enhancing classification capabilities, and adapting the scale of low-level and high-level features to facilitate fusion.

[0048] In order to better fuse feature information of different scales, the PANet structure is used for multi-scale feature fusion. The fusion formula is:

[0049]

[0050] Among them, F out It is the feature map output after feature fusion; F up It is a high-level feature map after upsampling operation, which contains more abstract and semantic features; F down It is the underlying feature map with higher resolution and richer detail information; Represents the feature splicing operation, F up and Fdown The splicing is performed in the channel dimension so that the fused feature map contains both high-level semantic information and low-level detail information.

[0051] In the YOLO Head structure, the output of the second five-level convolution module is connected to the input of the first detection head, the output of the third five-level convolution module is connected to the input of the second detection head, the output of the fourth five-level convolution module is connected to the input of the third detection head, and the output of the first detection head, the output of the second detection head, and the output of the third detection head are connected to the input of the output layer at the same time.

[0052] Each detection head has three prior boxes (anchors), which are generated by clustering the ground truth target boxes in the dataset using the K-means algorithm. When clustering noisy target boxes, nine cluster centers are randomly determined. Then, the intersection-of-union ratio between each noisy target box in the dataset and the cluster center is calculated, and the target box is assigned to the cluster center with the largest intersection-of-union ratio. After traversing the entire dataset, new cluster centers are recalculated based on the target box distribution until the cluster center no longer changes. The specific cluster target formula is shown below, where N represents the number of all noisy bounding boxes and K is the number of cluster centers.

[0053]

[0054] Where dist represents the value of the clustering objective function, which is used to measure the quality of the clustering effect. M represents the number of all noise bounding boxes in the dataset, which is the total number of samples traversed when calculating the clustering objective function. K is the pre-set number of cluster centers. i Refers to the i-th noise bounding box in the dataset. Each noise bounding box contains information such as the position and size of the noise in the image. j Represents the jth cluster center, which is an abstract representation of a certain type of noise feature. ∩ represents b i with c j The intersection operation calculates the area of the overlapping part of the two boxes. ∪ represents b i with c j The union operation of the two boxes is used to calculate the total area covered by the two boxes.

[0055] Because the original YOLOv4 prior bounds are derived from clustering of the COCO dataset and are suitable for detecting a variety of targets, they are not suitable for the noise detection task of this invention. Therefore, this invention recalculates nine prior bounds using the above method, with a minimum width and height of [8, 30] and a maximum width and height of [12, 70]. The aspect ratios of these prior bounds match the morphology of noisy targets, as noise distribution in an image differs from that of regular targets.

[0056] Given that noise detection is sensitive to detail information, increasing the downsampling factor will increase the receptive field of the feature map, resulting in the loss of noise details. Therefore, it is not suitable to use a detection head with a large downsampling factor for noise detection. Therefore, the present invention removes the 32x downsampling detection head in the original YOLOv4 model. However, considering that the noise detection task of the present invention requires high precision, otherwise it will affect the subsequent image processing and analysis, a 4x downsampling detection head with higher resolution is selected. To this end, the present invention ultimately replaces the original detection head with a downsampling factor of [8,16,32] with a detection head with a downsampling factor of [4,8,16].

[0057] The improved YOLOv4 retains its multi-scale detection advantages while significantly reducing the number of parameters by optimizing the network parameter configuration. Specifically, the k-means algorithm is used to regenerate the prior bounding box for the noise target, so that the size and proportion of the prior bounding box better reflect the actual distribution of noise in the image. Furthermore, based on the high requirement for detailed information in noise detection, the downsampling factor of the detection head is rationally adjusted. The inappropriate 32x downsampling detection head is removed and a 4x downsampling detection head is selected. The number of layers in the backbone network CSPDarknet53 is also reduced to obtain CSPDarknet48. These improvements significantly improve the network's operational efficiency while maintaining detection accuracy. The improved YOLOv4 network can quickly locate burst noise areas and dynamically assign weights based on the noise characteristics and distribution, giving higher attention weights to noisy areas, thereby achieving precise detection, location, and classification of noise.

[0058] Afterwards, the noise detection signal n3 is integrated into the CNN optimized signal, i.e., the noise suppression signal y1, as additional decision information in a point-by-point encoding manner, to form a four-dimensional feature vector signal containing amplitude and noise prior, i.e., the fused signal y2. Specifically:

[0059] First, expand the encoding dimension of noise category and position: expand the 3D one-hot encoding of the noise detection signal n3 (for example, Gaussian noise corresponds to [1,0,0], salt and pepper noise corresponds to [0,1,0], and composite noise corresponds to [0,0,1]) into a noise feature matrix of the same length as the noise suppression signal y1. The specific operation is: construct the expansion matrix C∈R N×4For the i-th row C(i,:) of the matrix C, c1, c2, and c3 are the one-hot encodings of the noise category (consistent with the category encoding of the noise detection signal n3), and c4 is the position encoding (if signal point i is within the actual noise location range, c4 is 1, otherwise it is 0). For example, if a certain interval is detected as Gaussian noise, then signal point i within that interval corresponds to C(i,:) = [1, 0, 0, 1], clearly indicating that it belongs to the Gaussian noise category and is located in the noise location; signal point j outside the interval corresponds to C(j,:) = [1, 0, 0, 0], indicating that it belongs to the Gaussian noise category but is not located in the noise location. This matrix not only implements the annotation of noise category information, but also reflects the local characteristics of the noise location through the position encoding c4, ensuring that each signal point carries prior information about the noise type and location.

[0060] Then, feature splicing and normalization are performed: the noise suppression signal y1 (dimension N×1, which is the voltage signal after CNN optimization, and each element corresponds to the pixel brightness quantization value) is spliced with the expanded noise feature matrix C in the channel dimension to obtain the fusion signal y2, y2=[y1,C]∈R N×4 After the splicing is completed, the feature of each signal point is expressed as [y 1,i ,c1,c2,c3,c4], realizing the point-by-point association of signal amplitude, noise category and position, and providing a more accurate multi-dimensional feature basis for subsequent processing.

[0061] Step 4: Use the SCAN decoding algorithm to decode the fused signal y2 to obtain the restored signal u1.

[0062] The SCAN decoding algorithm leverages the fused signal y2 to dynamically adjust SCAN decoding parameters based on noise classification results. The extracted noise distribution parameters are precisely embedded in the soft information iterative update mechanism of the polar code SCAN decoding. Through point-by-point log-likelihood ratio (LLR) calculation, soft information iteration, multi-path scanning, and optimal path selection, it achieves precise error correction for Gaussian, salt-and-pepper, and composite noise. This significantly reduces the bit error rate from that of the original signal, significantly improves decision confidence, and ultimately outputs the optimized restored signal u1. This transforms the grayscale histogram distribution in the image from dispersed to concentrated, significantly reducing edge blur, and effectively restoring noise-corrupted image details.

[0063] Application of fusion signal y2 in SCAN decoding:

[0064] 1) Log-likelihood ratio (LLR) dynamic correction The SCAN decoding module recalculates the LLR value based on the fused signal y2:

[0065]

[0066] Where LLR(i) represents the log-likelihood ratio at the i-th symbol position; P(b i =0|y3(i)) means sending bit b under the condition of fusing the i-th element of signal y2 i The probability of being 0; P(b i =1|y3(i)) means sending bit b under the condition of fusing the i-th element of signal y2 i The probability of 1.

[0067] The noise category codes c1, c2, and c3 are used as prior probability inputs to correct the assumptions about noise distribution in traditional LLR calculations (e.g., under the Gaussian noise assumption, the difference in the probability density functions of P(n|Gaussian) and P(n|Impulse) is explicitly introduced).

[0068] 2) Multi-path scanning strategy optimization. SCAN decoding dynamically adjusts the path retention strategy for different noise categories: 2 optimal paths are retained in Gaussian noise scenarios (to reduce computational complexity); 8 paths are retained in salt and pepper noise scenarios (to enhance burst error tolerance); and the number of adaptive paths N = 4 to 8 in composite noise scenarios (dynamically adjusted according to the noise mixing ratio output by YOLOv4).

[0069] In the calculation of path weights, the noise category encoding is used as an additional feature dimension to participate in the path metric update:

[0070]

[0071] Among them, Metric(path) is the metric value of the current decoding path, which is used to evaluate the quality of the current decoding path; y1(i) is the value of the voltage signal y1 at the i-th symbol position after CNN optimization; b i is the bit value of the i-th symbol position in the current decoding path, which can be 0 or 1; γ is the noise category penalty factor; c k Represents the elements in the noise category code (k=1, 2, 3), representing the encoding information of different noise categories (Gaussian, salt and pepper, composite noise), and participates in the path metric update; is the indicator function, indicating that when b i Typical error patterns corresponding to noise types When they are not equal, the function value is 1, otherwise it is 0; N is the polar code length.

[0072] During the polar code SCAN decoding process, the key decoding parameters are adaptively adjusted based on the noise estimation results of CNN and YOLOv4 to achieve refined processing of Gaussian, salt and pepper, and composite noise. While maintaining decoding efficiency, the bit error rate under complex interference is reduced by 40% to 50%, effectively improving the decoding algorithm's adaptability to noise. This method is significantly superior to the method of directly optimizing polar code decoding through deep learning.

[0073] It should be noted that although the embodiments of the present invention described above are illustrative, they are not intended to limit the present invention. Therefore, the present invention is not limited to the above-mentioned specific embodiments. Without departing from the principles of the present invention, any other embodiments obtained by those skilled in the art under the guidance of the present invention are deemed to be within the protection of the present invention.

Claims

1. A polar code image signal decoding method using CNN and YOLOv4 for indirect noise estimation, characterized by: The steps include: Step 1: pre-judge the received noisy signal y to generate a pre-judgement binary sequence s1, and subtract the pre-judgement binary sequence s1 from the noisy signal y to obtain a preliminary noise estimation signal n1; Step 2: The preliminary noise estimation signal n1 is fed into the improved CNN for noise feature extraction to obtain the noise feature signal n2, and the noise feature signal n2 is subtracted from the noisy signal y to obtain the noise suppression signal y1; Step 3: Send the noise suppression signal y1 to the improved YOLOv4 for noise detection to obtain the noise detection signal n3, and then concatenate the noise suppression signal y1 with the noise detection signal n3 to obtain the fusion signal y2; Step 4: Use the SCAN decoding algorithm to decode the fused signal y2 to obtain the restored signal u1.

2. The polar code image signal decoding method for indirect noise estimation using CNN and YOLOv4 according to claim 1, characterized in that: In step 2, the improved CNN consists of an input layer, four convolutional layers, four activation function layers, and an output layer; The input of the input layer serves as the input of the improved CNN; the output of the input layer is connected to the input of the first convolutional layer, the output of the first convolutional layer is connected to the input of the first activation function layer, the output of the first activation function layer is connected to the input of the second convolutional layer, the output of the second convolutional layer is connected to the input of the second activation function layer, the output of the second activation function layer is connected to the input of the third convolutional layer, the output of the third convolutional layer is connected to the input of the third activation function layer, the output of the third activation function layer is connected to the input of the fourth convolutional layer, the output of the fourth activation function layer is connected to the input of the output layer; the output of the output layer serves as the output of the improved CNN.

3. The polar code image signal decoding method for indirect noise estimation using CNN and YOLOv4 according to claim 1, characterized in that: In step 3, the improved YOLOv4 consists of an input layer, a two-dimensional convolutional layer, a batch normalization layer, an activation function layer, four residual modules, a spatial pyramid pooling module, two convolutional layers, two three-level convolutional modules, four five-level convolutional modules, two upsampling layers, two downsampling layers, four splicing layers, three detection heads, and an output layer; The input of the input layer serves as the input of the improved YOLOv4; the output of the input layer is connected to the input of the two-dimensional convolutional layer, the output of the two-dimensional convolutional layer is connected to the input of the batch normalization layer, the output of the batch normalization layer is connected to the input of the activation function layer, and the output of the activation function layer is connected to the input of the first residual module; the output of the first residual module is connected to the input of the second residual module, the output of the second residual module is connected to the input of the third residual module, the output of the third residual module is connected to the input of the fourth residual module, and the output of the fourth residual module is connected to the input of the first three-stage convolutional module; The output of the first three-level convolution module is connected to the input of the spatial pyramid pooling module, the output of the spatial pyramid pooling module is connected to the input of the second three-level convolution module, the output of the second three-level convolution module is connected to the input of the first convolution layer, the output of the first convolution layer is connected to the input of the first upsampling layer, the output of the first upsampling layer and the output of the third residual module are simultaneously connected to the input of the first splicing layer, the output of the first splicing layer is connected to the input of the first five-level convolution module, the output of the first five-level convolution module is connected to the input of the second convolution layer, the output of the second convolution layer is connected to the input of the second upsampling layer, and the output of the second upsampling layer is connected. The output of the second residual module is connected to the input of the second splicing layer at the same time, the output of the second splicing layer is connected to the input of the second five-level convolution module, the output of the second five-level convolution module is connected to the input of the first downsampling, the output of the first downsampling and the output of the first five-level convolution module are connected to the input of the third splicing layer at the same time, the output of the third splicing layer is connected to the input of the third five-level convolution module, the output of the third five-level convolution module is connected to the input of the second downsampling, the output of the second downsampling and the output of the second three-level convolution module are connected to the input of the fourth splicing layer at the same time, and the output of the fourth splicing layer is connected to the input of the fourth five-level convolution module; The output of the second five-level convolution module is connected to the input of the first detection head, the output of the third five-level convolution module is connected to the input of the second detection head, and the output of the fourth five-level convolution module is connected to the input of the third detection head; the output of the first detection head, the output of the second detection head, and the output of the third detection head are simultaneously connected to the input of the output layer; the output of the output layer serves as the output of the improved YOLOv4.

4. The polar code image signal decoding method for indirect noise estimation using CNN and YOLOv4 according to claim 3, wherein: The first residual module includes 1 residual module, the second residual module and the third residual module each include 8 repeated residual modules, and the fourth residual module includes 4 repeated residual modules.

Images