Search

CN-121981912-A - Frequency modulation and wavelet sub-band guided dual-domain collaborative transform X-ray image denoising method

CN121981912ACN 121981912 ACN121981912 ACN 121981912ACN-121981912-A

Abstract

The invention discloses a frequency modulation and wavelet sub-band guided double-domain collaborative transform X-ray image denoising method, which comprises the steps of collecting a noisy digital ray original image and a corresponding clear reference image thereof, constructing a data set after preprocessing, constructing a network model of a double-domain collaborative coding-decoding architecture, wherein shallow layer characteristics of an input image are extracted through 3X 3 depth convolution, encoding stages alternately stack ETB and AFMB (analog to digital) to represent local and global information, a WB-LKED (noise-free block) module is embedded to strengthen fine granularity characteristics, WDB performs downsampling and transmits high frequency characteristics to a decoding end, a decoding stage WUB restores resolution through double-way upsampling, a coder is fused with the same-scale characteristics and utilizes the high frequency characteristics to strengthen details, the data set is refined through the ETB and the AFMB after the characteristics are spliced, an output characteristic is obtained through 3X 3 depth convolution, and a global residual error connection optimization result is combined, and a training model is utilized to input an image to be denoised to output a final result. The invention can overcome the problems of weak anti-complex noise interference capability, poor detail retaining effect and limited CNR improvement.

Inventors

  • LIU LINGLING
  • HU MENGHUI
  • WU LINGFENG
  • AO BO

Assignees

  • 南昌航空大学

Dates

Publication Date
20260505
Application Date
20260403

Claims (10)

  1. 1. The frequency modulation and wavelet sub-band guided double-domain collaborative transducer X-ray image denoising method is suitable for weld DR image denoising and ceramic matrix composite CT slice denoising, and is characterized by comprising the following steps: Collecting a digital ray original image containing noise and a corresponding clear reference image thereof, and preprocessing the digital ray original image to construct a data set; The method comprises the steps of constructing a network model of a frequency modulation and wavelet sub-band guided double-domain collaborative coding-decoding architecture, alternately stacking ETB and AFMB (electronic toll collection) in a coding stage to jointly represent local and global information, embedding a WB-LKED module to strengthen fine granularity characteristic representation, performing downsampling by WDB and transmitting high-frequency characteristics to a decoding end through high-frequency connection, recovering resolution by WUB through double-way upsampling, adaptively fusing the same-scale characteristics with the encoder, utilizing high-frequency characteristic enhancement details, splicing the fused characteristics with corresponding level characteristics of the encoder, refining the fused characteristics with the AFMB step by step, obtaining output characteristics through 3X 3 depth convolution of an output end, and adding the input characteristics with the output characteristics through global residual connection; and training the network model by utilizing the data set, inputting an image to be denoised, and outputting a final denoising result.
  2. 2. The frequency modulation and wavelet sub-band guided double-domain collaborative transform X-ray image denoising method as claimed in claim 1, wherein a DR imaging experiment platform is built in an image acquisition process according to NB/T47013.11-2015 standard, a multi-angle double-wall double-shadow transillumination mode is adopted for a small-diameter tube test piece, a single vertical transillumination mode is adopted for a flat plate lap joint weld test piece, a digital ray original image containing noise and a corresponding clear reference image under the same imaging angle are obtained by adjusting imaging parameters, and acquired image sensitivity, resolution, gray scale range and normalized signal to noise ratio meet the requirements of AB class detection technology.
  3. 3. The method for denoising the frequency modulation and wavelet sub-band guided dual-domain collaborative transform X-ray Image according to claim 1, wherein the preprocessing comprises the steps of carrying out window width and window level adjustment on an acquired noise-containing digital ray original Image and a corresponding clear reference Image thereof by adopting Image J software and deriving the noise-containing digital ray original Image into a JPG format Image, carrying out region clipping on the JPG format Image, reserving effective detection regions of welding lines and a heat affected zone, and removing ineffective background regions.
  4. 4. The frequency modulated and wavelet sub-band guided dual domain cooperative transform X-ray image denoising method of claim 1, wherein said ETB comprises MHTA modules and GDFN modules; The MHTA module performs layer normalization on input features, extracts local space features and cross-channel information through 1×1 point-by-point convolution and 3×3 depth convolution in sequence, obtains query vectors, key vectors and value vectors through channel partitioning, performs feature rearrangement, calculates correlation between the query vectors and the key vectors through matrix multiplication, obtains channel attention weights through a Softmax activation function, performs weighted aggregation on the value vectors, and obtains output through 1×1 convolution and residual addition; The GDFN module performs layer normalization on MHTA module output again, sequentially divides the features into two paths through 1×1 convolution, 3×3 depth convolution and channel blocking, wherein one path is processed by GELU activation function, the other path is used as gating signal, and the two paths of results are multiplied element by element, then are subjected to 1×1 convolution and added with input residual errors to be output.
  5. 5. A frequency modulated and wavelet sub-band guided dual domain collaborative transform X-ray image denoising process as claimed in claim 1, wherein the AFMB comprises an FM phase and an FI phase; In the FM stage, 3X 3 depth convolution is carried out on a degraded image to align with an input feature, the degraded image is mapped to a frequency domain through FFT to obtain a frequency spectrum, an MGB and LFMB are used for adaptively generating a high-frequency mask and a low-frequency mask, the MGB sequentially adopts GAP, 1X 1 convolution, GELU activation function, 1X 1 convolution and Sigmoid activation function to generate a space constraint coefficient, LFMB generates a low-frequency mask with a space center area of 1 and an edge area of 0, the high-frequency mask is obtained by inverting the low-frequency mask, the high-frequency mask and the low-frequency mask are multiplied by the frequency spectrum element by element respectively and restored to a high-frequency feature and a low-frequency feature through IFFT, and frequency decoupling is realized; In the FI stage, high-frequency characteristics and low-frequency characteristics are interacted in a cross-frequency manner through an SA module and a CA module, the SA module respectively executes GAP and GMP on the high-frequency characteristics in a channel dimension to obtain two space description graphs, space attention force diagram is generated through 7X 7 depth separable convolution and Sigmoid activation function after splicing, space fusion characteristics are obtained through element-by-element multiplication with the low-frequency characteristics, the CA module respectively executes GAP and GMP on the low-frequency characteristics in the space dimension to obtain two groups of channel descriptors, channel attention weight is generated through two layers of 1X 1 convolution and ReLU activation function transformation and then is added and fused, channel fusion characteristics are obtained through channel attention weight multiplication with the high-frequency characteristics through the Sigmoid activation function, the space fusion characteristics and the channel fusion characteristics are spliced in the channel dimension and are obtained through 1X 1 convolution, and the fusion characteristics are input to the MHTCA module to execute cross attention, so that high-low-frequency characteristic self-adaptive modulation is achieved, and final output of the AFMB is obtained.
  6. 6. The method of denoising a frequency modulated and wavelet sub-band guided dual domain collaborative transform X-ray image of claim 5, wherein said MHTCA module processing comprises: the method comprises the steps of taking frequency domain features and input features as double inputs, sequentially carrying out 1×1 convolution and 3×3 depth convolution on the frequency domain features to generate query vectors, and sequentially carrying out 1×1 convolution, 3×3 depth convolution and channel blocking on the input features to obtain key vectors and value vectors; After feature rearrangement is carried out on the query vector, the key vector and the value vector, the relevance of the query vector and the key vector is calculated, attention weight is obtained through Softmax activation, the value vector is weighted and summed, multi-head fusion is completed through residual connection and 1X 1 convolution, and the frequency-space cross attention enhancement feature is generated.
  7. 7. The frequency modulated and wavelet sub-band guided dual domain cooperative transform X-ray image denoising method of claim 1, wherein said WDB processing comprises: The input features are decomposed into a low-frequency feature and three high-frequency features by adopting a DWT, and the four sub-band features are spliced in the channel dimension and subjected to 1X 1 convolution projection to obtain compression features; Adding two high-frequency features with low-frequency features element by element through a wavelet attention module, adding the two paths of added results through 3X 3 depth convolution aggregation field context, and generating a first attention map through 1X 1 convolution and Sigmoid activation after the two paths of aggregation features are added again; The first attention is multiplied and weighted with the compressed characteristic element by element, residual connection is carried out with the compressed characteristic to obtain the output characteristic of WDB, and simultaneously, three high-frequency characteristics are transmitted to a WUB module at a decoding end through high-frequency connection for subsequent IDWT operation.
  8. 8. The method for denoising a frequency modulated and wavelet sub-band guided dual domain collaborative transform X-ray image of claim 1, wherein said WUB process comprises: Based on the high-frequency characteristics transferred by WDB, the two-way up-sampling is realized by combining IDWT and pixel rearrangement, weighted by a CAA module and refined by a fusion mobile inverse bottleneck convolution module; The two-way sampling is concretely characterized in that a first branch carries out channel mapping on input features through 1 multiplied by 1 convolution and realizes spatial resolution up-sampling through pixel rearrangement, a second branch inputs high-frequency features transmitted by WDB and the input features into IDWT together, high-resolution detail recovery is realized by utilizing high-frequency detail information, and two-way sampling results are spliced in channel dimension to obtain pre-fusion features; The CAA module calculates the local mean value and the global standard deviation of the channel for the pre-fusion feature, adds the local mean value and the global standard deviation, and generates a second attention map through 1X 1 convolution, a ReLU activation function, 1X 1 convolution and a Sigmoid activation function in sequence, and multiplies the second attention map by the pre-fusion feature to obtain a weighted feature; And splitting and adding the weighted features according to channels to obtain fusion features, refining the fusion features through a fusion mobile inverse bottleneck convolution module, wherein the fusion mobile inverse bottleneck convolution module sequentially comprises 1×1 convolution, 3×3 group convolution, siLU activation functions and 1×1 convolution, and finally outputting up-sampling features.
  9. 9. A frequency modulated and wavelet sub-band guided dual domain cooperative transform X-ray image denoising method as defined in claim 1, wherein said WB-LKED module is formed by a cascade of WB and LKED modules; the method comprises the steps that WB carries out DWT decomposition on input features to obtain low-frequency features and three groups of high-frequency features, the low-frequency features are refined by FB module and then fused with the three groups of high-frequency features by IDWT to restore resolution, the FB module comprises a two-layer channel confusion shuffler and a one-layer fusion mobile inverse bottleneck convolution module, the channel confusion shuffler consists of channel projection, 7×7 depth convolution and another layer channel projection, the channel projection is realized through layer normalization, channel blocking, point-by-point MLP, identity mapping, channel splicing and channel shuffling, and gradient stability is maintained through residual connection; The LKED module sequentially carries out three-stage treatment of distillation, condensation and enhancement, wherein the distillation stage adopts three stages of distillation units to be stacked in series, each distillation unit is formed by a channel screening compression submodule and a BSRB characteristic refinement submodule containing 3X 3 depth convolution in parallel, and the 3X 3 depth convolution is added after the final stage distillation unit; And adding the output of the LKED module with the input feature through the global residual connection to obtain a WB-LKED final output feature.
  10. 10. The method of claim 1, wherein the model training uses a batch size of 1, an optimizer of AdamW, a momentum parameter of β 1 =0.9,β 2 =0.999, a numerical stability term of ε=1× -8 , a loss function of L1, a cosine annealing learning rate scheduler combining a preheating mechanism and a restarting mechanism, an initial learning rate of 2× -4 , a minimum learning rate of 1× -6 , an input image size of 256×256 pixels, and training rounds of 60 rounds.

Description

Frequency modulation and wavelet sub-band guided dual-domain collaborative transform X-ray image denoising method Technical Field The invention relates to the technical field of image processing and ray digital imaging detection, in particular to a frequency modulation and wavelet sub-band guided double-domain collaborative transform X-ray image denoising method. Background The radiation detection is used as one of the core means of industrial nondestructive detection, is widely applied to the key fields of aerospace, nuclear industry, special equipment and the like, has the core value of accurately identifying internal defects of materials and components, and provides important technical support for structural safety guarantee and quality tracing. However, due to the synergistic constraint of the detector and readout electronics, imaging link stability, exposure conditions (tube voltage, tube current, exposure time) and workpiece scattering, the actually acquired X-ray image often superimposes quantum fluctuation noise and electronic readout noise simultaneously, forming a typical poisson-gaussian mixture noise. The noise can directly cause the image contrast to be reduced and the signal to noise ratio to be reduced, so that the visibility of the micro defects is greatly attenuated, the difficulty of evaluating the artificial defects is increased, the defects can be missed to be detected and misjudged, and the accuracy and the reliability of industrial nondestructive detection are seriously affected. Therefore, the method is oriented to the actual scene of industrial ray detection, develops a high-efficiency image denoising technology, becomes a core link for improving the quality of an X-ray image, enhancing the visibility of defects, improving the detection rate of the defects and evaluating the reliability, and designs a high-efficiency model capable of balancing between noise reduction and detail reservation, and is more a key for solving the problems of noise interference and detail loss. At present, X-ray image denoising research forms two main technical paths, namely a traditional method taking statistical modeling and transform domain processing as cores, and a deep learning method relying on implicit characterization of big data learning. The research of the traditional denoising method focuses on explicit noise prior construction and transform domain feature extraction, various classical filtering strategies and models are derived, and a certain denoising effect can be achieved in a simple noise scene. But is limited by the technical principle, the traditional method has extremely strong dependence on noise priori, and needs to manually adjust a large number of parameters, when facing complex composite scattering, space non-uniform noise in industrial scenes, or X-ray images with abundant structural details, the denoising performance of the method can be obviously reduced, and the balance of noise suppression and detail reservation is difficult to be considered. Along with the accumulation of large-scale image data sets and the improvement of computing hardware performance, the X-ray image denoising method based on deep learning is rapidly developed, and becomes a hot spot direction of current research. The prior researches have proposed a plurality of denoising schemes based on CNN (Convolutional Neural Network ) and Transformer, GAN (GENERATIVE ADVERSARIAL Network, generating an countermeasure Network), and by introducing the technologies of a filtering module, a attention mechanism, adaptive learning and the like, remarkable progress is made in noise suppression and key detail preservation, and great application potential of deep learning in the field of image denoising is fully shown, but the core pain point of unbalanced noise reduction and detail preservation is still not completely solved. Although the deep learning denoising algorithm breaks through part of the limitations of the traditional method, an obvious short plate still exists in an industrial X-ray image denoising scene, wherein the existing deep learning model is easy to generate an excessive smooth phenomenon in the denoising process, is difficult to accurately distinguish Noise from tiny defect details, causes degradation of image details and insufficient CNR (Contrast-to-Noise Ratio), and cannot meet the strict requirements of industrial nondestructive detection on accurate defect identification. In conclusion, the existing method has three general core technical problems of weak anti-complex noise interference capability, poor detail retaining effect and limited CNR improvement. Disclosure of Invention The invention provides a frequency modulation and wavelet sub-band guided double-domain collaborative transform X-ray image denoising method, which solves the three core technical problems of weak anti-complex noise interference capability, poor detail retaining effect and limited CNR improvement existing in the existing metho