Industrial nondestructive testing image denoising method based on feature domain defect perception

Abstract

The invention discloses an industrial nondestructive testing image denoising method based on feature domain defect perception, and the method innovatively fuses a defect perception mechanism and deep learning denoising to solve the problem of missing detection caused by easy smoothing of tiny defect features when an industrial image is processed by the existing denoising technology. The method is characterized in that a brand-new U-Net denoising algorithm model is designed, an attention mechanism is introduced, a novel defect perception loss function is introduced, a semi-supervised joint training strategy is combined, a denoising network is guided to effectively suppress background noise, key tiny defect features are identified and reserved, and the defect perception loss function is improved. Therefore, the defect detection rate of the denoised image is obviously improved.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing technology, specifically relating to an industrial non-destructive testing image denoising method based on feature domain defect perception. Background Technology

[0002] In the field of industrial nondestructive testing, imaging technologies such as X-rays and ultrasound are widely used for the inspection of internal product structures and defect identification. However, due to factors such as inherent noise of imaging equipment, environmental interference, and the complexity of material structures, the acquired images often contain a large amount of noise, which seriously affects the accuracy of defect identification and detection. Especially for critical defects such as microcracks, pores, and inclusions, traditional noise reduction methods are prone to smoothing or losing defect features while filtering out noise, leading to missed detections and misjudgments, posing potential risks to industrial quality control and safe production.

[0003] Traditional image denoising methods can be mainly divided into filter-based spatial domain methods and transform-domain-based frequency domain methods. While these methods can suppress noise to some extent, they often rely on manually designed prior assumptions and are poorly adaptable to complex and variable industrial image noise models. They also struggle to achieve an effective balance between noise suppression and feature preservation. Especially in situations with low signal-to-noise ratios, traditional methods easily lead to blurred image edges and loss of detail, failing to meet the stringent requirements of high-precision non-destructive testing for defect feature preservation.

[0004] With the development of deep learning technology, diffusion models have emerged. These models gradually add noise to data and then train a model to learn how to gradually reverse this process. For example, XIAO ZHISHENG's patent "DENOISINGDIFFUSION GENERATIVE ADVERSARIAL NETWORKS" uses a diffusion model and proposes a method combining denoising diffusion with generative adversarial networks. However, this method adds noise based on noise in natural images, and the noise inevitably overlaps. In the field of industrial non-destructive testing, noise has special characteristics. Adding noise requires extracting various types of noise first, which is not easy to implement in practice. Noise superposition can also obscure certain key feature defects. Directly applying this to non-destructive testing images often improves the appearance quality of the image but weakens the sensitivity to defect areas, making it difficult to meet the core requirement of "no missed defects" in industrial inspection.

[0005] Leveraging the advantages of deep learning and frequency-division attention mechanisms, a novel defect-aware loss function is introduced, and combined with a semi-supervised joint training strategy, an efficient denoising framework is constructed. This method effectively guides the denoising network to identify and enhance the feature representation of key minute defects while suppressing complex background noise, thereby achieving accurate preservation of defect structures during the denoising process. Summary of the Invention

[0006] The purpose of this invention is to address the problem that traditional denoising methods easily cause the smoothing or loss of defect features while filtering out noise, leading to missed detections and misjudgments. This invention provides an industrial non-destructive testing image denoising method based on feature domain defect perception.

[0007] To solve the above problems, the present invention is achieved through the following technical solution:

[0008] A method for denoising industrial nondestructive testing images based on feature domain defect perception includes the following steps:

[0009] Step 1: Prepare the dataset, which consists of three parts. One part is used as the training set for model training, another part is used as the validation set in the model training nodes, and the third part is used as the test set for testing the performance of the network model.

[0010] Step 2: Design a brand new U-Net denoising algorithm model.

[0011] Furthermore, in step two, a frequency division attention module is introduced into the new U-Net denoising algorithm model.

[0012] Step 3: Introduce a novel defect-aware loss function.

[0013] Step 4: Adopt a semi-supervised joint training strategy, which integrates supervised pre-training and generative adversarial fine-tuning.

[0014] Furthermore, in step one, the dataset uses publicly available industrial datasets. Image preprocessing is performed on the dataset, and the industrial images are normalized using the following formula:

[0015]

[0016] in, For the original image, and These are the minimum and maximum pixel values ​​of the image, respectively.

[0017] Furthermore, in step two, a novel U-Net denoising algorithm model is constructed, embedding a frequency division attention module before each hop connection in the encoder and decoder. Its core operation can be represented as:

[0018]

[0019] in, and These represent the attention weights for high-frequency and low-frequency components, respectively. This is element-wise multiplication.

[0020] To address the overfitting problem in supervised learning caused by excessively deep network models and numerous parameters, this invention employs a variant of the Dropout function, defined as follows:

[0021] For mask variables that follow a Bernoulli distribution, Preserve the probability for the neuron. To maintain the desired scaling factor for the output, its calculation formula is as follows:

[0022]

[0023] Furthermore, in step three, a defect-aware joint loss function is designed. This function is composed of noise suppression loss. Loss with defects Common components:

[0024]

[0025] in, Denoising images constrained by smooth L1 loss function With real and clear images Overall consistency; Through defect region mask Compared to real defect masks The grayscale difference loss between the two is used to enhance the preservation of local features in the defect area, and the coefficient is... , Used to balance the contributions of the two losses.

[0026] Furthermore, in step four, supervised noise modeling pre-training is performed first to obtain a preliminary denoising basic model.

[0027] In this study, a Generative Adversarial Network (GAN) is introduced. The generator model adopts a novel U-Net denoising algorithm model, and the discriminator adopts the PatchGAN architecture. It performs independent forward propagation on the input image and finally outputs an N×N matrix as the discrimination result. Each element in the matrix represents the confidence of the corresponding image patch. This result is then fed into the loss function to drive the adversarial training of the generator and the discriminator.

[0028] The adversarial loss function between the generator and the discriminator is as follows:

[0029] Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention.

[0031] Figure 2 This is a schematic diagram of the U-Net denoising algorithm model structure of the present invention.

[0032] Figure 3 A schematic diagram of the generative adversarial network structure of this invention.

[0033] Figure 4 Comparison of denoised images for an industrial dataset.

[0034] Figure 5 A comparison of images denoised and feature-preserving for industrial datasets.

[0035] Figure 6 This is a comparison table of test metrics based on an industrial dataset. Detailed Implementation

[0036] The present invention will be further illustrated below with reference to specific embodiments.

[0037] Step 1: Collect a large number of industrial non-destructive testing images, totaling 1000, including both defective and defect-free images, as the total dataset.

[0038] The data was partitioned by randomly dividing the total dataset into training, validation, and test sets in a 7:2:1 ratio. The training set was used for model parameter learning, the validation set was used for hyperparameter tuning and monitoring the training process, and the test set was used for objective evaluation of the final model performance.

[0039] Further data augmentation is performed. To improve the model's generalization ability, online data augmentation strategies are used on the training set images, including random horizontal / vertical flipping, small-angle random rotation (±10°), and minor perturbations to brightness and contrast. The corresponding defect label images also need to be synchronized.

[0040] Image preprocessing: Before inputting the image into the network, all images undergo pixel value normalization, linearly transforming their range from [0, 255] to [0, 1]. The normalization formula is as follows:

[0041]

[0042] Step 2: Design a novel U-Net denoising algorithm model. The encoder part uses four downsampling stages, each containing two 3×3 convolutional layers. The Batch Normalization (BN) layer within the convolutional layers introduces a variant of the Dropout function. Each convolutional layer is followed by a ReLU activation function and a 2×2 max-pooling layer for downsampling. The decoder part also contains four upsampling stages, each using transposed convolutions for upsampling and concatenating the upsampling with the feature maps of the corresponding layers in the encoder (Skip-Connection), followed by two 3×3 convolutional layers.

[0043] The Frequency Domain Attention Module (FDAM) is added before each hop connection in the encoder and decoder, and its characteristics are as follows: for the input feature map Global context information is extracted using two branches: global average pooling and global max pooling. The outputs of these two branches are then fed into a shared small multilayer perceptron (MLP) to generate channel attention weight vectors. The sigmoid function is used to normalize the weight vectors to the range [0,1] to obtain the final feature weights. , weight Compared with the original input feature map Element-wise multiplication enables adaptive feature calibration. This module effectively suppresses feature responses corresponding to background noise while enhancing the weights of high-frequency features associated with minor defects.

[0044] The loss function used is a defect-aware loss function, consisting of two weighted components, designed to simultaneously optimize global denoising and local defect preservation. The calculation formula is as follows:

[0045]

[0046] in and To balance the hyperparameters, in this embodiment, the values ​​were experimentally verified to be 1.0 and 0.8, respectively.

[0047] The training process employs a semi-supervised learning framework, with the first stage being supervised training. The input for this stage is a noisy image. The new U-Net denoising algorithm model outputs a clear image. The network was trained using a defect-aware loss function to learn how to process noisy images. To clear image The mapping.

[0048] The second stage is unsupervised training, which uses a Generative Adversarial Network (GAN). The generator (G) is the U-Net denoising algorithm model pre-trained in the first stage, and its input is the noisy image. The output is a clear image. .

[0049] The discriminator (D) employs a PatchGAN structure, consisting of 5 convolutional layers, ultimately outputting an N×N matrix (rather than a single scalar). Each element in the matrix represents the probability that a local region (patch) in the image is a real image. This structure better captures the local details and textures of the image, forcing the generator to produce realistic outputs even in local regions.

[0050] Extract a batch of defective noisy images-clear images-defect mask triples from the training set. , , ).

[0051] Generator receives Generate denoised image .

[0052] and real, clear images The samples are fed into discriminator D to obtain their discrimination probability maps. and .

[0053] All outputs are used to calculate the total loss and update the parameters of the generator and discriminator.

[0054] The training in this phase is guided by a multi-objective loss function, with the generator (G) experiencing a total loss. as follows.

[0055] The generator aims to simultaneously fool the discriminator, maintain overall image fidelity, and accurately preserve defects. Its total loss consists of three parts:

[0056]

[0057] Noise suppression loss Similar to the first stage, the reconstruction accuracy of the entire image domain is constrained.

[0058] Defect retention loss This loss function only applies to real defect masks. The effect applies within the marked area. Specifically, it calculates the pixel values ​​of the defective areas in the denoised image. Pixel values ​​of defect areas in a real, clear image The L1 distance between them. This directly penalizes any smoothing or blurring operations the network performs on defective regions, strongly guiding it to preserve the contrast and shape of the defects. The calculation formula is:

[0059] Combat losses The calculation formula is:

[0060] The goal of generator G is to maximize the probability that discriminator D classifies its output as "true". .

[0061] The goal of the discriminator is a binary classification task: correctly distinguish between real and sharp images. Denoising images generated by the generator The calculation formula is as follows: The discriminator D needs to maximize the accuracy of judging true images. The probability of false positives in the generated image is maximized. The probability of.

[0062] At this point, training employs an alternating iterative approach, fixing the generator G and updating the discriminator D: calculating... Backpropagation updates the parameters of D (multiple iterations).

[0063] With discriminator D fixed, generator G is updated: calculation Backpropagation updates the parameters of G (1 iteration).

[0064] Configure key hyperparameters, and set the loss weights as follows: , , Higher This ensures the core objective of defect retention, with smaller Prevent excessive interference from combat losses with the main mission.

[0065] Both the generator and discriminator use the Adam optimizer. The generator's initial learning rate is set to 1×10⁻⁶. −4 The initial learning rate of the discriminator is set to 5×10. −5 .

[0066] The learning rate employs a cosine annealing strategy, gradually decreasing the learning rate during training to help the model converge to a better local optimum.

[0067] Train the model and save the weight model file with the lowest loss value.

[0068] Finally, the industrial noise image is denoised.

[0069] This invention uses peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) to measure the overall denoising effect, and uses signal retention rate (SSR) to measure the retention of defect features.

[0070] During the test, PSNR, SSIM, and SSR were calculated and recorded for the denoised image and 100 labeled images, respectively.

[0071] This invention provides an image denoising method for industrial non-destructive testing based on feature domain defect perception, guiding the denoising network to effectively suppress background noise while identifying and preserving key minute defect features. It should be noted that the embodiments described in this invention are for illustrative purposes only and do not constitute a limitation on the scope of protection of this invention. Therefore, this invention is not limited to the specific embodiments described above. Any other embodiments derived by those skilled in the art through reasonable deduction or association under the guidance of the principles of this invention should be considered to fall within the scope of protection of this invention.

Claims

1. A method for denoising industrial nondestructive testing images based on feature domain defect perception, characterized in that, We designed a novel U-Net denoising algorithm model, introduced an attention mechanism, a novel defect-aware loss function, and a semi-supervised joint training strategy, which effectively suppressed background noise while retaining key minute defect features.

2. The method according to claim 1, characterized in that, A brand-new U-Net denoising algorithm model.

3. The method according to claim 2, characterized in that, By introducing a frequency division attention mechanism module, feature information can be extracted more effectively.

4. The method according to claim 1, characterized in that, The defect perception loss function consists of noise suppression loss and defect retention loss.

5. The method according to claim 1, characterized in that, The semi-supervised joint training strategy consists of supervised pre-training and a generative adversarial network.

6. The semi-supervised joint training strategy according to claim 5, characterized in that, Generative Adversarial Networks (GANs) consist of two sub-networks: a generator and a discriminator. The generator model adopts a novel U-Net denoising algorithm model, while the discriminator uses the PatchGAN architecture.

Images