Adaptive perception-based inverse-consistent printed image registration method

Abstract

This invention discloses an inverse consistency printed image registration method based on adaptive perception. First, a printed image registration dataset is constructed, and the images are normalized. A deep learning network is then constructed for non-rigid registration of printed images. Next, the forward deformation field from the reference image to the image to be registered is predicted. This forward deformation field is input into a local adaptive perception module to obtain an optimized deformation field that highlights the true local nonlinear deformation. Finally, the registration network is optimized end-to-end based on a joint loss function, and the optimized deformation field is used to perform a spatial transformation on the printed image to be registered, outputting a registration result aligned with the reference printed image. This invention achieves accurate modeling of local nonlinear deformation in printed images, thereby improving the accuracy and stability of printed image registration and providing a reliable foundation for subsequent high-precision defect detection tasks.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and industrial image processing technology, specifically relating to an inverse consistency printed image registration method based on adaptive perception. Background Technology

[0002] In the automated quality inspection of printed materials, reference-based defect detection methods are widely used. The core premise of these methods is that the reference printed image and the image of the printed material to be inspected are precisely aligned at the pixel level. However, in actual image acquisition, factors such as differences in the flexibility of the printing material, external forces, flatness of placement, and imaging angle often cause localized nonlinear deformations in the image of the printed material to be inspected, such as localized stretching, twisting, or wrinkling.

[0003] The aforementioned localized non-rigid deformations can lead to a large number of spurious difference regions in traditional defect detection methods based on pixel difference or feature matching, significantly reducing the accuracy and stability of defect detection. Therefore, before defect detection, it is usually necessary to geometrically align the reference printed image and the printed image to be registered using image registration techniques.

[0004] Existing traditional image registration methods often rely on manually designed features or complex iterative optimization processes. When faced with complex local nonlinear deformations, they are not only computationally inefficient but also have limited deformation modeling capabilities. In recent years, unsupervised image registration methods based on deep learning have been gradually applied to nonlinear deformation registration tasks. However, existing methods generally suffer from several problems: most methods only model unidirectional deformation fields, lacking constraints on the consistency of forward and reverse registration results; in local nonlinear deformation scenarios, most areas in the image do not actually deform, making it difficult for existing methods to effectively distinguish between truly deformed and undeformed areas, easily introducing meaningless deformation estimates. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive perception-based inverse consistency printed image registration method, which enables accurate modeling of local nonlinear deformations in printed images, thereby improving the accuracy and stability of printed image registration and providing a reliable foundation for subsequent high-precision defect detection tasks.

[0006] The technical solution adopted in this invention is an adaptive perception-based inverse consistency printed image registration method, which is implemented according to the following steps: Step 1: Perform local nonlinear deformation enhancement processing on the acquired printed images to construct a printed image registration dataset, and perform uniform size normalization processing on the images to meet the input requirements of the registration network. Step 2: Construct a deep learning network for non-rigid registration of printed images. The network is based on an encoder-decoder structure with shared parameters to predict dense deformation fields. Inverse consistency deformation constraints and adaptive perception mechanisms are introduced into the network to enhance the consistency and local discrimination ability of deformation estimation. Step 3: Input the reference printed image and the printed image to be registered into the deep learning network to predict the forward deformation field from the reference image to the image to be registered; at the same time, predict the corresponding reverse deformation field by swapping the image input order, and apply inverse consistency constraints to the forward and reverse deformation fields so that they satisfy the inverse relationship in the sense of spatial mapping.

[0007] Step 4: Input the positive deformation field into the local adaptive sensing module composed of a multilayer perceptron (MLP), judge the pixel-level deformation state, and adaptively adjust the deformation field according to the judgment result. Keep the displacement information unchanged in the area determined to be deformed, and weaken the displacement response of the area determined to be undeformed, thereby obtaining an optimized deformation field that highlights the real local nonlinear deformation. Step 5: Perform end-to-end optimization of the registration network based on the joint loss function, which includes at least image similarity constraints, deformation field smoothing constraints, inverse consistency constraints, and adaptive perception constraints; and use the optimized deformation field to perform spatial transformation on the printed image to be registered, outputting a registration result aligned with the reference printed image.

[0008] The invention is further characterized in that, Step 1 is implemented in the following steps: Training and testing datasets were constructed. Local nonlinear data augmentation techniques were applied to synthesize and augment the collected printed image dataset to construct a printed image registration dataset. The images were uniformly scaled to 192×160 to meet the network input size requirements. The entire dataset contains 25,200 data pairs, of which the training set contains 20,160 image pairs and the testing set contains 5,040 image pairs.

[0009] Step 1 is as follows: Step 1.1: Collect images of various types of printed materials on the laboratory machine vision platform to form a basic image set. During the acquisition, ensure that the printed content of the reference sample is clear, the exposure and contrast are usable, and record the local deformation phenomenon under typical non-ideal working conditions for the design of enhancement strategies. Step 1.2: Employ a local nonlinear deformation enhancement strategy to construct small-area local distortion, stretching, or wrinkling phenomena, and construct a printed image dataset for registration. The reference printed image is denoted as [image name missing]. Generate the image to be registered as follows: : Step 1.2.1, in Randomly crop a candidate local rectangular region. ; Step 1.2.2, for The direction of the inner pixel is random and the amplitude is limited. Random perturbations within a certain range are used to obtain the boundary of the local region after the perturbation, and the perturbation region is constrained not to exceed the boundary; if it exceeds the boundary, it is resampled and generated. Step 1.2.3: Apply the above local disturbance to get This results in the formation of registered sample pairs with local nonlinear deformation. ; Step 1.2.4: Repeat this process. In this way, each reference image can generate multiple images to be registered for different deformation instances, thus obtaining the final dataset used for training. Step 1.3: Size normalization and input normalization. The images are uniformly cropped to the 192×160 size required by the network input and grayscale normalization is performed to meet the tensor size consistency requirements of the network batch training.

[0010] Step 2 is implemented in the following steps: A deep learning network for print image registration, incorporating an inverse consistency deformation strategy and an adaptive perception mechanism, is constructed. The specific structure of the deep learning network is as follows: (1) Shared parameter dual-branch U-Net registration backbone network: This backbone network includes a feature encoder and a feature decoder with skip connections. The network adopts a dual-branch structure, in which the forward branch Source images in ascending order and target image As input, the features are extracted by the encoder through downsampling, and then fused with the decoder through skip connections to output a forward dense deformation field. Simultaneously, the input order of the images is swapped in reverse order. and As input, the reverse branch, which shares network parameters with the forward branch, is fed in. In the middle, the output is a reverse dense deformation field. ; (2) Local adaptive sensing module: This module is connected to the output of the forward branch and is composed of a multilayer sensing mechanism, with a dense forward deformation field. The data are fed into a multilayer perceptron for feature mapping, which predicts the perception results used to distinguish between truly deformed and undeformed regions. Subsequently, the fine-tuning function module receives the sensing result. With the initial deformation field After fine-tuning, a more accurate high-precision deformation field is output. ; (3) Spatial transformation network: This network includes a grid sampler connected after the fine-tuning function module to receive the fine-tuned high-precision deformation field. and source image As input, the source image is processed using a deformation field mesh. The pixel space is subjected to a differentiable resampling transformation, and the final output is the registered deformed image. ; (4) Multiple loss function constraint mechanism: In the end-to-end connection of the network, multiple loss functions are set to optimize the network weights; specifically, real labels are used at the output of the multilayer perceptron. right Calculate the loss of the perceptual label In fine-tuning deformation field Calculate smoothing loss To ensure the spatial continuity of the deformation field; the registered image is calculated at the network output. With target image loss of similarity between ; and in the forward deformation field With reverse deformation field Apply inverse consistency loss between To improve the consistency and symmetry of deformation estimation under different input orders.

[0011] Step 2 is as follows: Step 2.1: Refer to the printed image Image of the printed matter to be registered As input to the registration deep learning network, the output of the network's forward computation is a two-dimensional dense deformation field mapped from the reference image coordinate system to the coordinate system of the image to be registered, denoted as... For any pixel This represents the predicted two-dimensional displacement vector at that point, where, Indicates in the deformation field Under its influence, specific pixels in the image The deformation displacement vector generated at the location; Represents pixels The displacement generated in the horizontal direction of the two-dimensional coordinate system after network prediction; Represents pixels The displacement generated in the vertical direction of the two-dimensional coordinate system after network prediction; Step 2.2: The network adopts a U-Net encoder-decoder structure. The encoder downsamples stepwise to extract multi-scale representations to simultaneously cover local texture details and a wider range of structural semantics. The decoder upsamples stepwise to restore spatial resolution and fuses features from the corresponding scales of the encoder through skip connections, thus balancing localization accuracy and deformation continuity when outputting the deformation field. Finally, a dense deformation field with the same resolution as the input is regressed. .

[0012] In step 3, the reference printed image will be used. Image of the printed matter to be registered The data is input into a registration deep learning network with shared parameters, and then processed through convolutional feature extraction and decoding regression to obtain the data from... arrive positive deformation field Then swap the input order of the two images, and... Input the inverse uniform deformation network to predict the reverse deformation field During training, an inverse consistency loss is constructed to ensure that the forward and reverse deformation fields satisfy inverse constraints in the sense of spatial mapping, i.e., by minimizing deformation synthesis. The square norm is made to approach zero, which improves the symmetry and stability of deformation estimation and reduces the registration error caused by changes in input order.

[0013] Step 3 is implemented in the following steps: Step 3.1: Obtain the reference printed image to be registered. Image of the printed matter to be registered The two are then input into a registration deep learning network with shared parameters in a predetermined order, and the output is a positive deformation field. ,in Represents the pixel points in the reference image coordinate system The displacement vector mapped to the coordinate system of the image to be registered; Step 3.2, Exchange and The input order will A registration deep learning network with the same shared parameters outputs an inverse deformation field. ,in Represents the pixel points in the coordinate system of the image to be registered. The displacement vector when mapped back to the reference image coordinate system; Step 3.3: Based on the positive deformation field Calculate pixels Corresponding continuous coordinates in the image to be registered The coordinates u satisfy:

[0014] in The coordinates are non-integer coordinates; Step 3.4: For continuous coordinates Perform neighborhood interpolation sampling to obtain In position Displacement value at; determine The set of integer grid points in the neighborhood For any neighborhood point Calculate the bilinear interpolation weights:

[0015] in and They represent spatial dimensions. The coordinate components on; Step 3.5, define the resultant displacement of the forward and reverse deformation fields as:

[0016] The symbol " " indicates the deformation field synthesis operation; when and When the transformations are inverses, the above synthesis result should approach the zero vector; Step 3.6: Construct the inverse consistency loss function For the domain The composite displacement of all pixels within the area is constrained by the L2 norm to minimize the composite result. The loss function is:

[0017] in It is a norm 2; Step 3.7: During network training, the inverse consistency loss is... As a component of the joint loss function, it participates in backpropagation to simultaneously constrain... and The inverse relationship between the inputs is used to maintain the consistency of the registration results when the input order is changed, and to improve the symmetry and stability of the deformation field estimation.

[0018] In step 4, the positive deformation field is... The input is fed into the adaptive sensing module, where a multilayer perceptron determines the pixel-level deformation state and outputs the deformation region prediction result. ,in accordance with right Adaptive fine-tuning is performed: for regions determined to have undergone deformation, the displacement direction and amplitude information are kept unchanged; for regions determined not to have undergone significant deformation, the displacement intensity is attenuated, weakening its direction and amplitude response, in order to suppress ineffective or redundant deformation and highlight the real local nonlinear deformation regions, thus obtaining the fine-tuned deformation field. .

[0019] Step 4 is implemented in the following steps: Step 4.1: Obtain the positive deformation field output in Step 3. ,in Represents the pixel points in the reference image coordinate system Two-dimensional displacement vector; As input features for the local adaptive sensing module; Step 4.2: The deformation field... Perform pixel-level feature organization, dividing each pixel point The displacement vector component at that location is used as the input feature of that pixel. The data is then input pixel-by-pixel into a multilayer perceptron for deformation state determination; wherein the multilayer perceptron contains at least one hidden layer and a nonlinear activation function, and the linear transformation of its neurons satisfies:

[0020] in As weight, For bias; Step 4.3: Output deformation discrimination results for each pixel using a multilayer perceptron. ,in Indicates whether the pixel belongs to a local deformation region; when When representing "deformation", the pixel point Included in the set of deformation regions ;when When indicating "undeformed", the pixel is... Included in the set of non-deformable regions ; Step 4.4: Deformation judgment results For positive deformation field Perform pixel-level fine-tuning and construct the fine-tuning function. To optimize the deformation field The fine-tuning function satisfies: when the pixel point When a region is identified as a deformation region, the displacement vector remains unchanged; when a pixel... When a region is determined to be non-deformable, the displacement vector is attenuated to weaken invalid deformation. The fine-tuning function expression is as follows:

[0021] in The attenuation coefficient; Step 4.5: Output the optimized deformation field obtained after fine-tuning. , The original displacement direction and amplitude information are preserved in the deformation region, and the displacement response is suppressed in the non-deformation region, so that the deformation field exhibits more significant local deformation distribution differences in space, thereby improving the ability to capture local nonlinear deformation and reducing the interference of redundant deformation on the registration results. Step 4.6: During the network training phase, deformation label loss is used to constrain the local adaptive perception module: obtain the deformation labels corresponding to the image to be registered. And the deformation labels are matched with the prediction results. Calculate label loss To ensure that the deformation region discrimination result output by the multilayer perceptron is consistent with the deformation label; the label loss is preferably in the form of mean square error.

[0022] in This represents the number of pixels or the number of samples included in the statistics.

[0023] Step 5 is implemented in the following steps: The image registration dataset for printed materials generated through local nonlinear data augmentation is divided into training and testing datasets. Network input samples are constructed by pairing reference images with images to be registered. During network training, the model parameters are iteratively updated based on the principle of stochastic gradient descent to gradually minimize the joint loss function composed of multiple constraints. The joint loss function consists of image similarity loss, deformation field smoothing regularization loss, inverse consistency loss, and deformation region label loss. Each loss term participates in end-to-end backpropagation optimization through weighted combination to collaboratively constrain the learning and generation of the optimal deformation field.

[0024] The beneficial effects of this invention are: (1) The inverse consistency printed image registration method based on adaptive perception can effectively model and accurately align the nonlinear deformations such as local stretching and twisting in printed images without the need for real deformation field annotation, significantly reducing the false differences caused by imaging errors or material deformation, and providing a stable and reliable spatial alignment basis for subsequent differential-based printed defect detection and quality assessment tasks. (2) By introducing inverse consistency deformation constraints in the registration process, this invention ensures that the forward deformation field and the reverse deformation field from the reference image to the image to be registered satisfy a reciprocal relationship in the sense of spatial mapping, effectively suppressing the instability problem of deformation estimation caused by changes in input order or unidirectional registration bias, and enhancing the symmetry and physical rationality of the deformation field. (3) This invention further introduces a local adaptive perception mechanism to distinguish the deformation state of different spatial regions in the deformation field, and makes differentiated adjustments to the deformation field based on the discrimination results, so that the region where real nonlinear deformation occurs can fully retain its displacement direction and amplitude information, while effectively suppressing invalid or redundant displacement responses in the region where no obvious deformation occurs, thereby avoiding the problem of global excessive deformation and highlighting the local real deformation features. (4) Under a unified end-to-end unsupervised learning framework, this invention organically integrates image similarity constraints, deformation field smoothing constraints, inverse consistency constraints and deformation region perception constraints, so that the deformation field learning process simultaneously meets the requirements of geometric continuity, spatial consistency and local deformation sensitivity, and achieves efficient and accurate registration of non-rigid deformation of complex printed materials under the premise of ensuring controllable computational complexity. Attached Figure Description

[0025] Figure 1 This is a flowchart of the inverse consistency printed image registration method based on adaptive perception of the present invention; Figure 2(a) is a partial sample diagram of the dataset for the image registration task provided in Embodiment 6 of the present invention; Figure 2(b) is a partial sample diagram of the dataset for the image registration task provided in Embodiment 6 of the present invention; Figure 3 This is a network architecture diagram of an inverse consistency printed image registration method based on adaptive perception provided in Embodiment 6 of the present invention; Figure 4 This is a schematic diagram of the inverse consistency deformation strategy in an inverse consistency printed image registration method based on adaptive perception provided in Embodiment 6 of the present invention. Figure 5 This is a schematic diagram of the MLP structure in the local adaptive perception mechanism of the inverse consistency printed image registration method based on adaptive perception provided in Embodiment 6 of the present invention. Figure 6The image to be registered, the reference image, the registration deformation field, and the registration result are provided in Embodiment 6 of the present invention. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0027] This invention relates to an adaptive sensing-based inverse consistency printed image registration method, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Perform local nonlinear deformation enhancement processing on the acquired printed images to construct a printed image registration dataset, and perform uniform size normalization processing on the images to meet the input requirements of the registration network. Part of the dataset is shown in Figure 2(a) and Figure 2(b). Step 1 is implemented in the following steps: Training and testing datasets were constructed, and local nonlinear data augmentation techniques were applied to synthesize and augment the collected printed image datasets. A printed image registration dataset containing non-rigid deformation features such as local stretching and twisting was constructed. The images were uniformly scaled to 192×160 to meet the network input size requirements. The entire dataset contains 25,200 data pairs, of which the training set contains 20,160 image pairs and the test set contains 5,040 image pairs. Some results are shown in Figure 2(a) and Figure 2(b).

[0028] Step 1 is as follows: Step 1.1: Collect images of various types of printed materials, such as medicine boxes, cigarette and alcohol boxes, and food packaging boxes, on a laboratory machine vision platform to form a basic image set. During the acquisition, ensure that the printed content of the reference sample is clear, the exposure and contrast are usable, and record the local deformation phenomena under typical non-ideal working conditions for the design of enhancement strategies. Step 1.2: Employ a local nonlinear deformation enhancement strategy to construct small-area local distortion, stretching, or wrinkling phenomena, and construct a printed image dataset for registration. The reference printed image is denoted as [image name missing]. Generate the image to be registered as follows: : Step 1.2.1, in Randomly crop a candidate local rectangular region. ; Step 1.2.2, for The direction of the inner pixel is random and the amplitude is limited. Random perturbations within a certain range are used to obtain the boundary of the local region after the perturbation, and the perturbation region is constrained not to exceed the boundary; if it exceeds the boundary, it is resampled and generated. Step 1.2.3: Apply the above local disturbance to get This results in the formation of registered sample pairs with local nonlinear deformation. ; Step 1.2.4: Repeat this process. In this way, each reference image can generate multiple images to be registered for different deformation instances, resulting in the final dataset used for training; a portion of the dataset is shown in Figure 2.

[0029] Step 1.3: Size normalization and input normalization. The images are uniformly cropped to the 192×160 size required by the network input and grayscale normalization is performed to meet the tensor size consistency requirements of the network batch training.

[0030] Step 2: Construct a deep learning network for non-rigid registration of printed images, such as... Figure 3 As shown, the network is based on an encoder-decoder structure with shared parameters for predicting dense deformation fields; and inverse consistency deformation constraints and adaptive sensing mechanisms are introduced into the network to enhance the consistency and local discrimination capability of deformation estimation. Step 2 is implemented in the following steps: Construct a deep learning network for print image registration that incorporates an inverse consistency deformation strategy and an adaptive perception mechanism, such as... Figure 3 As shown, the specific structure of a deep learning network is as follows: (1) Shared parameter dual-branch U-Net registration backbone network: This backbone network includes a feature encoder and a feature decoder with skip connections. The network adopts a dual-branch structure, in which the forward branch Source images in ascending order and target image As input, the features are extracted by the encoder through downsampling, and then fused with the decoder through skip connections to output a forward dense deformation field. Simultaneously, the input order of the images is swapped in reverse order. and As input, the reverse branch, which shares network parameters with the forward branch, is fed in. In the middle, the output is a reverse dense deformation field. ; (2) Local adaptive sensing module: This module is connected to the output of the forward branch and is composed of a multilayer sensing mechanism, with a dense forward deformation field. The data are fed into a multilayer perceptron for feature mapping, which predicts the perception results used to distinguish between truly deformed and undeformed regions. Subsequently, the fine-tuning function module receives the sensing result. With the initial deformation field After fine-tuning, a more accurate high-precision deformation field is output. ; (3) Spatial transformation network: This network includes a grid sampler connected after the fine-tuning function module to receive the fine-tuned high-precision deformation field. and source image As input, the source image is processed using a deformation field mesh. The pixel space is subjected to a differentiable resampling transformation, and the final output is the registered deformed image. ; (4) Multiple loss function constraint mechanism: In the end-to-end connection of the network, multiple loss functions are set to optimize the network weights; specifically, real labels are used at the output of the multilayer perceptron. right Calculate the loss of the perceptual label In fine-tuning deformation field Calculate smoothing loss To ensure the spatial continuity of the deformation field; the registered image is calculated at the network output. With target image loss of similarity between ; and in the forward deformation field With reverse deformation field Apply inverse consistency loss between To improve the consistency and symmetry of deformation estimation under different input orders.

[0031] Step 2 is as follows: Step 2.1: Refer to the printed image Image of the printed matter to be registered As input to the registration deep learning network, the output of the network's forward computation is a two-dimensional dense deformation field mapped from the reference image coordinate system to the coordinate system of the image to be registered, denoted as... For any pixel This represents the predicted two-dimensional displacement vector at that point, where, Indicates in the deformation field Under its influence, specific pixels in the image The deformation displacement vector generated at the location; Represents pixels The displacement generated in the horizontal direction of the two-dimensional coordinate system after network prediction; Represents pixels The displacement generated in the vertical direction of the two-dimensional coordinate system after network prediction; Step 2.2: The network adopts a U-Net encoder-decoder structure, such as... Figure 3As shown, the encoder progressively downsamples to extract multi-scale representations, simultaneously covering local texture details and broader structural semantics; the decoder progressively upsamples to restore spatial resolution and fuses features from the corresponding scales of the encoder through skip connections, thus balancing localization accuracy and deformation continuity when outputting the deformation field; finally, it regresses a dense deformation field with the same resolution as the input. .

[0032] Step 3: Input the reference printed image and the printed image to be registered into the deep learning network to predict the forward deformation field from the reference image to the image to be registered; simultaneously, predict the corresponding reverse deformation field by swapping the image input order, and apply an inverse consistency constraint to the forward and reverse deformation fields to ensure that they satisfy a reciprocal relationship in the sense of spatial mapping, such as... Figure 4 As shown.

[0033] In step 3, the reference printed image will be used. Image of the printed matter to be registered The data is input into a registration deep learning network with shared parameters, and then processed through convolutional feature extraction and decoding regression to obtain the data from... arrive positive deformation field Then swap the input order of the two images, and... Input the inverse uniform deformation network to predict the reverse deformation field ,like Figure 4 As shown, an inverse consistency loss is constructed during training to ensure that the forward and reverse deformation fields satisfy the inverse constraint in the sense of spatial mapping, i.e., by minimizing deformation synthesis. The square norm is made to approach zero, which improves the symmetry and stability of deformation estimation and reduces the registration error caused by changes in input order.

[0034] Step 3 is implemented in the following steps: Step 3.1: Obtain the reference printed image to be registered. Image of the printed matter to be registered The two are then input into a registration deep learning network with shared parameters in a predetermined order, and the output is a positive deformation field. ,in Represents the pixel points in the reference image coordinate system The displacement vector mapped to the coordinate system of the image to be registered; Step 3.2, Exchange and The input order will A registration deep learning network with the same shared parameters outputs an inverse deformation field. ,in Represents the pixel points in the coordinate system of the image to be registered. The displacement vector when mapped back to the reference image coordinate system; Step 3.3: Based on the positive deformation field Calculate pixels Corresponding continuous coordinates in the image to be registered The coordinates u satisfy:

[0035] in The coordinates are non-integer coordinates; Step 3.4: For continuous coordinates Perform neighborhood interpolation sampling to obtain In position Displacement value at; determine The set of integer grid points in the neighborhood For any neighborhood point Calculate the bilinear interpolation weights:

[0036] in and They represent spatial dimensions. The coordinate components on; Step 3.5, define the composite displacement (inverse uniform composite) of the forward and reverse deformation fields as:

[0037] The symbol " " indicates the deformation field synthesis operation; when and When the transformations are inverses, the above synthesis result should approach the zero vector; Step 3.6: Construct the inverse consistency loss function For the domain The composite displacement of all pixels within the area is constrained by the L2 norm to minimize the composite result. The loss function is:

[0038] in It is a norm 2; Step 3.7: During network training, the inverse consistency loss is... As a component of the joint loss function, it participates in backpropagation to simultaneously constrain... and The inverse relationship between the inputs is used to maintain the consistency of the registration results when the input order is changed, and to improve the symmetry and stability of the deformation field estimation.

[0039] Step 4: Input the positive deformation field into the local adaptive sensing module composed of a multilayer perceptron (MLP), determine the pixel-level deformation state, and adaptively adjust the deformation field based on the determination result, such as... Figure 5 As shown, the displacement information of the region determined to be deformed is kept unchanged, while the displacement response of the region determined not to be deformed is weakened, thereby obtaining an optimized deformation field that highlights the real local nonlinear deformation. In step 4, such as Figure 5 As shown, the positive deformation field The input is fed into the adaptive sensing module, where a multilayer perceptron determines the pixel-level deformation state and outputs the deformation region prediction result. ,in accordance with right Adaptive fine-tuning is performed: for regions determined to have undergone deformation, the displacement direction and amplitude information are kept unchanged; for regions determined not to have undergone significant deformation, the displacement intensity is attenuated, weakening its direction and amplitude response, in order to suppress ineffective or redundant deformation and highlight the real local nonlinear deformation regions, thus obtaining the fine-tuned deformation field. .

[0040] Step 4 is implemented in the following steps: Step 4.1: Obtain the positive deformation field output in Step 3. ,in Represents the pixel points in the reference image coordinate system Two-dimensional displacement vector; As input features for the local adaptive sensing module; Step 4.2: The deformation field... Perform pixel-level feature organization, dividing each pixel point The displacement vector component at that location is used as the input feature of that pixel. The data is then input pixel-by-pixel into a multilayer perceptron for deformation state determination; wherein the multilayer perceptron contains at least one hidden layer and a nonlinear activation function, and the linear transformation of its neurons satisfies:

[0041] in As weight, For bias; Step 4.3: Output deformation discrimination results for each pixel using a multilayer perceptron. ,in Indicates whether the pixel belongs to a local deformation region; when When representing "deformation", the pixel point Included in the set of deformation regions ;when When indicating "undeformed", the pixel is... Included in the set of non-deformable regions ; Step 4.4: Deformation judgment results For positive deformation field Perform pixel-level fine-tuning and construct the fine-tuning function. To optimize the deformation field The fine-tuning function satisfies: when the pixel point When a region is identified as a deformation region, the displacement vector remains unchanged; when a pixel... When a region is determined to be non-deformable, the displacement vector is attenuated to weaken invalid deformation. The fine-tuning function expression is as follows:

[0042] in The attenuation coefficient is denoted by ; the preferred value is . .

[0043] Step 4.5: Output the optimized deformation field obtained after fine-tuning. , The original displacement direction and amplitude information are preserved in the deformation region, and the displacement response is suppressed in the non-deformation region, so that the deformation field exhibits more significant local deformation distribution differences in space, thereby improving the ability to capture local nonlinear deformation and reducing the interference of redundant deformation on the registration results. Step 4.6: During the network training phase, deformation label loss is used to constrain the local adaptive perception module: obtain the deformation labels corresponding to the image to be registered. And the deformation labels are matched with the prediction results. Calculate label loss To ensure that the deformation region discrimination result output by the multilayer perceptron is consistent with the deformation label; the label loss is preferably in the form of mean square error.

[0044] in This represents the number of pixels or the number of samples included in the statistics.

[0045] Step 5: Perform end-to-end optimization of the registration network based on the joint loss function, which includes at least image similarity constraints, deformation field smoothing constraints, inverse consistency constraints, and adaptive perception constraints; and use the optimized deformation field to perform spatial transformation on the printed image to be registered, outputting a registration result aligned with the reference printed image.

[0046] Step 5 is implemented in the following steps: The image registration dataset for printed materials generated through local nonlinear data augmentation was divided into training and testing datasets. Network input samples were constructed by pairing reference and target printed images. During network training, model parameters were iteratively updated based on stochastic gradient descent to gradually minimize the joint loss function, which consists of multiple constraints. This joint loss function comprises image similarity loss, deformation field smoothing regularization loss, inverse consistency loss, and deformation region label loss. These loss terms are weighted and combined to participate in end-to-end backpropagation optimization, collaboratively constraining the learning and generation of the optimal deformation field. The network input image size was set to (192, 160), the training batch size to 32, and the initial learning rate to 1×10⁻⁶. -4 The number of training iterations was set to 150; the Adam optimization algorithm based on adaptive first-moment estimation and second-moment estimation was used to implement the stochastic gradient descent update strategy to improve the convergence speed and stability of the training process.

[0047] Example 1 This invention relates to an adaptive sensing-based inverse consistency printed image registration method, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Perform local nonlinear deformation enhancement processing on the acquired printed images to construct a printed image registration dataset, and perform uniform size normalization processing on the images to meet the input requirements of the registration network. Part of the dataset is shown in Figure 2. Step 2: Construct a deep learning network for non-rigid registration of printed images, such as... Figure 3 As shown, the network is based on an encoder-decoder structure with shared parameters for predicting dense deformation fields; and inverse consistency deformation constraints and adaptive sensing mechanisms are introduced into the network to enhance the consistency and local discrimination capability of deformation estimation. Step 3: Input the reference printed image and the printed image to be registered into the deep learning network to predict the forward deformation field from the reference image to the image to be registered; simultaneously, predict the corresponding reverse deformation field by swapping the image input order, and apply an inverse consistency constraint to the forward and reverse deformation fields to ensure that they satisfy a reciprocal relationship in the sense of spatial mapping, such as... Figure 4 As shown.

[0048] Step 4: Input the positive deformation field into the local adaptive sensing module composed of a multilayer perceptron (MLP), determine the pixel-level deformation state, and adaptively adjust the deformation field based on the determination result, such as... Figure 5 As shown, the displacement information of the region determined to be deformed is kept unchanged, while the displacement response of the region determined not to be deformed is weakened, thereby obtaining an optimized deformation field that highlights the real local nonlinear deformation. Step 5: Perform end-to-end optimization of the registration network based on the joint loss function, which includes at least image similarity constraints, deformation field smoothing constraints, inverse consistency constraints, and adaptive perception constraints; and use the optimized deformation field to perform spatial transformation on the printed image to be registered, outputting a registration result aligned with the reference printed image.

[0049] Example 2 An adaptive perception-based inverse consistency printed image registration method, flowchart as follows: Figure 1 As shown, please follow these steps: Step 1: Perform local nonlinear deformation enhancement processing on the acquired printed images to construct a printed image registration dataset, and perform uniform size normalization processing on the images to meet the input requirements of the registration network. Part of the dataset is shown in Figure 2. Step 1 is implemented in the following steps: Training and testing datasets were constructed, and local nonlinear data augmentation techniques were applied to synthesize and augment the collected printed image dataset. A printed image registration dataset containing non-rigid deformation features such as local stretching and twisting was constructed. The images were uniformly scaled to 192×160 to meet the network input size requirements. The entire dataset contains 25,200 data pairs, of which the training set contains 20,160 image pairs and the testing set contains 5,040 image pairs. Some results are shown in Figure 2.

[0050] Step 1 is as follows: Step 1.1: Collect images of various types of printed materials, such as medicine boxes, cigarette and alcohol boxes, and food packaging boxes, on a laboratory machine vision platform to form a basic image set. During the acquisition, ensure that the printed content of the reference sample is clear, the exposure and contrast are usable, and record the local deformation phenomena under typical non-ideal working conditions for the design of enhancement strategies. Step 1.2: Employ a local nonlinear deformation enhancement strategy to construct small-area local distortion, stretching, or wrinkling phenomena, and construct a printed image dataset for registration. The reference printed image is denoted as [image name missing]. Generate the image to be registered as follows: : Step 1.2.1, in Randomly crop a candidate local rectangular region. ; Step 1.2.2, for The direction of the inner pixel is random and the amplitude is limited. Random perturbations within a certain range are used to obtain the boundary of the local region after the perturbation, and the perturbation region is constrained not to exceed the boundary; if it exceeds the boundary, it is resampled and generated. Step 1.2.3: Apply the above local disturbance to get This results in the formation of registered sample pairs with local nonlinear deformation. ; Step 1.2.4: Repeat this process. In this way, each reference image can generate multiple images to be registered for different deformation instances, resulting in the final dataset used for training; a portion of the dataset is shown in Figure 2.

[0051] Step 1.3: Size normalization and input normalization. The images are uniformly cropped to the 192×160 size required by the network input and grayscale normalization is performed to meet the tensor size consistency requirements of the network batch training.

[0052] Step 2: Construct a deep learning network for non-rigid registration of printed images, such as... Figure 3 As shown, the network is based on an encoder-decoder structure with shared parameters for predicting dense deformation fields; and inverse consistency deformation constraints and adaptive sensing mechanisms are introduced into the network to enhance the consistency and local discrimination capability of deformation estimation. Step 3: Input the reference printed image and the printed image to be registered into the deep learning network to predict the forward deformation field from the reference image to the image to be registered; simultaneously, predict the corresponding reverse deformation field by swapping the image input order, and apply an inverse consistency constraint to the forward and reverse deformation fields to ensure that they satisfy a reciprocal relationship in the sense of spatial mapping, such as... Figure 4 As shown.

[0053] Step 4: Input the positive deformation field into the local adaptive sensing module composed of a multilayer perceptron (MLP), determine the pixel-level deformation state, and adaptively adjust the deformation field based on the determination result, such as... Figure 5 As shown, the displacement information of the region determined to be deformed is kept unchanged, while the displacement response of the region determined not to be deformed is weakened, thereby obtaining an optimized deformation field that highlights the real local nonlinear deformation. Step 5: Perform end-to-end optimization of the registration network based on the joint loss function, which includes at least image similarity constraints, deformation field smoothing constraints, inverse consistency constraints, and adaptive perception constraints; and use the optimized deformation field to perform spatial transformation on the printed image to be registered, outputting a registration result aligned with the reference printed image.

[0054] Example 3 This invention relates to an adaptive sensing-based inverse consistency printed image registration method, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Perform local nonlinear deformation enhancement processing on the acquired printed images to construct a printed image registration dataset, and perform uniform size normalization processing on the images to meet the input requirements of the registration network. Part of the dataset is shown in Figure 2. Step 1 is implemented in the following steps: Training and testing datasets were constructed, and local nonlinear data augmentation techniques were applied to synthesize and augment the collected printed image dataset. A printed image registration dataset containing non-rigid deformation features such as local stretching and twisting was constructed. The images were uniformly scaled to 192×160 to meet the network input size requirements. The entire dataset contains 25,200 data pairs, of which the training set contains 20,160 image pairs and the testing set contains 5,040 image pairs. Some results are shown in Figure 2.

[0055] Step 1 is as follows: Step 1.1: Collect images of various types of printed materials, such as medicine boxes, cigarette and alcohol boxes, and food packaging boxes, on a laboratory machine vision platform to form a basic image set. During the acquisition, ensure that the printed content of the reference sample is clear, the exposure and contrast are usable, and record the local deformation phenomena under typical non-ideal working conditions for the design of enhancement strategies. Step 1.2: Employ a local nonlinear deformation enhancement strategy to construct small-area local distortion, stretching, or wrinkling phenomena, and construct a printed image dataset for registration. The reference printed image is denoted as [image name missing]. Generate the image to be registered as follows: : Step 1.2.1, in Randomly crop a candidate local rectangular region. ; Step 1.2.2, for The direction of the inner pixel is random and the amplitude is limited. Random perturbations within a certain range are used to obtain the boundary of the local region after the perturbation, and the perturbation region is constrained not to exceed the boundary; if it exceeds the boundary, it is resampled and generated. Step 1.2.3: Apply the above local disturbance to get This results in the formation of registered sample pairs with local nonlinear deformation. ; Step 1.2.4: Repeat this process. In this way, each reference image can generate multiple images to be registered for different deformation instances, resulting in the final dataset used for training; a portion of the dataset is shown in Figure 2.

[0056] Step 1.3: Size normalization and input normalization. The images are uniformly cropped to the 192×160 size required by the network input and grayscale normalization is performed to meet the tensor size consistency requirements of the network batch training.

[0057] Step 2: Construct a deep learning network for non-rigid registration of printed images, such as... Figure 3As shown, the network is based on an encoder-decoder structure with shared parameters for predicting dense deformation fields; and inverse consistency deformation constraints and adaptive sensing mechanisms are introduced into the network to enhance the consistency and local discrimination capability of deformation estimation. Step 2 is implemented in the following steps: Construct a deep learning network for print image registration that incorporates an inverse consistency deformation strategy and an adaptive perception mechanism, such as... Figure 3 As shown, the specific structure of a deep learning network is as follows: (1) Shared parameter dual-branch U-Net registration backbone network: This backbone network includes a feature encoder and a feature decoder with skip connections. The network adopts a dual-branch structure, in which the forward branch Source images in ascending order and target image As input, the features are extracted by the encoder through downsampling, and then fused with the decoder through skip connections to output a forward dense deformation field. Simultaneously, the input order of the images is swapped in reverse order. and As input, the reverse branch, which shares network parameters with the forward branch, is fed in. In the middle, the output is a reverse dense deformation field. ; (2) Local adaptive sensing module: This module is connected to the output of the forward branch and is composed of a multilayer sensing mechanism, with a dense forward deformation field. The data are fed into a multilayer perceptron for feature mapping, which predicts the perception results used to distinguish between truly deformed and undeformed regions. Subsequently, the fine-tuning function module receives the sensing result. With the initial deformation field After fine-tuning, a more accurate high-precision deformation field is output. ; (3) Spatial transformation network: This network includes a grid sampler connected after the fine-tuning function module to receive the fine-tuned high-precision deformation field. and source image As input, the source image is processed using a deformation field mesh. The pixel space is subjected to a differentiable resampling transformation, and the final output is the registered deformed image. ; (4) Multiple loss function constraint mechanism: In the end-to-end connection of the network, multiple loss functions are set to optimize the network weights; specifically, real labels are used at the output of the multilayer perceptron. right Calculate the loss of the perceptual label In fine-tuning deformation field Calculate smoothing loss To ensure the spatial continuity of the deformation field; the registered image is calculated at the network output. With target image loss of similarity between ; and in the forward deformation field With reverse deformation field Apply inverse consistency loss between To improve the consistency and symmetry of deformation estimation under different input orders.

[0058] Step 2 is as follows: Step 2.1: Refer to the printed image Image of the printed matter to be registered As input to the registration deep learning network, the output of the network's forward computation is a two-dimensional dense deformation field mapped from the reference image coordinate system to the coordinate system of the image to be registered, denoted as... For any pixel This represents the predicted two-dimensional displacement vector at that point, where, Indicates in the deformation field Under its influence, specific pixels in the image The deformation displacement vector generated at the location; Represents pixels The displacement generated in the horizontal direction of the two-dimensional coordinate system after network prediction; Represents pixels The displacement generated in the vertical direction of the two-dimensional coordinate system after network prediction; Step 2.2: The network adopts a U-Net encoder-decoder structure, such as... Figure 3 As shown, the encoder progressively downsamples to extract multi-scale representations, simultaneously covering local texture details and broader structural semantics; the decoder progressively upsamples to restore spatial resolution and fuses features from the corresponding scales of the encoder through skip connections, thus balancing localization accuracy and deformation continuity when outputting the deformation field; finally, it regresses a dense deformation field with the same resolution as the input. .

[0059] Step 3: Input the reference printed image and the printed image to be registered into the deep learning network to predict the forward deformation field from the reference image to the image to be registered; simultaneously, predict the corresponding reverse deformation field by swapping the image input order, and apply an inverse consistency constraint to the forward and reverse deformation fields to ensure that they satisfy a reciprocal relationship in the sense of spatial mapping, such as... Figure 4 As shown.

[0060] Step 4: Input the positive deformation field into the local adaptive sensing module composed of a multilayer perceptron (MLP), determine the pixel-level deformation state, and adaptively adjust the deformation field based on the determination result, such as... Figure 5As shown, the displacement information of the region determined to be deformed is kept unchanged, while the displacement response of the region determined not to be deformed is weakened, thereby obtaining an optimized deformation field that highlights the real local nonlinear deformation. Step 5: Perform end-to-end optimization of the registration network based on the joint loss function, which includes at least image similarity constraints, deformation field smoothing constraints, inverse consistency constraints, and adaptive perception constraints; and use the optimized deformation field to perform spatial transformation on the printed image to be registered, outputting a registration result aligned with the reference printed image.

[0061] Example 4 This invention relates to an adaptive sensing-based inverse consistency printed image registration method, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Perform local nonlinear deformation enhancement processing on the acquired printed images to construct a printed image registration dataset, and perform uniform size normalization processing on the images to meet the input requirements of the registration network. Part of the dataset is shown in Figure 2. Step 1 is implemented in the following steps: Training and testing datasets were constructed, and local nonlinear data augmentation techniques were applied to synthesize and augment the collected printed image dataset. A printed image registration dataset containing non-rigid deformation features such as local stretching and twisting was constructed. The images were uniformly scaled to 192×160 to meet the network input size requirements. The entire dataset contains 25,200 data pairs, of which the training set contains 20,160 image pairs and the testing set contains 5,040 image pairs. Some results are shown in Figure 2.

[0062] Step 1 is as follows: Step 1.1: Collect images of various types of printed materials, such as medicine boxes, cigarette and alcohol boxes, and food packaging boxes, on a laboratory machine vision platform to form a basic image set. During the acquisition, ensure that the printed content of the reference sample is clear, the exposure and contrast are usable, and record the local deformation phenomena under typical non-ideal working conditions for the design of enhancement strategies. Step 1.2: Employ a local nonlinear deformation enhancement strategy to construct small-area local distortion, stretching, or wrinkling phenomena, and construct a printed image dataset for registration. The reference printed image is denoted as [image name missing]. Generate the image to be registered as follows: : Step 1.2.1, in Randomly crop a candidate local rectangular region. ; Step 1.2.2, for The direction of the inner pixel is random and the amplitude is limited. Random perturbations within a certain range are used to obtain the boundary of the local region after the perturbation, and the perturbation region is constrained not to exceed the boundary; if it exceeds the boundary, it is resampled and generated. Step 1.2.3: Apply the above local disturbance to get This results in the formation of registered sample pairs with local nonlinear deformation. ; Step 1.2.4: Repeat this process. In this way, each reference image can generate multiple images to be registered for different deformation instances, resulting in the final dataset used for training; a portion of the dataset is shown in Figure 2.

[0063] Step 1.3: Size normalization and input normalization. The images are uniformly cropped to the 192×160 size required by the network input and grayscale normalization is performed to meet the tensor size consistency requirements of the network batch training.

[0064] Step 2: Construct a deep learning network for non-rigid registration of printed images, such as... Figure 3 As shown, the network is based on an encoder-decoder structure with shared parameters for predicting dense deformation fields; and inverse consistency deformation constraints and adaptive sensing mechanisms are introduced into the network to enhance the consistency and local discrimination capability of deformation estimation. Step 2 is implemented in the following steps: Construct a deep learning network for print image registration that incorporates an inverse consistency deformation strategy and an adaptive perception mechanism, such as... Figure 3 As shown, the specific structure of a deep learning network is as follows: (1) Shared parameter dual-branch U-Net registration backbone network: This backbone network includes a feature encoder and a feature decoder with skip connections. The network adopts a dual-branch structure, in which the forward branch Source images in ascending order and target image As input, the features are extracted by the encoder through downsampling, and then fused with the decoder through skip connections to output a forward dense deformation field. Simultaneously, the input order of the images is swapped in reverse order. and As input, the reverse branch, which shares network parameters with the forward branch, is fed in. In the middle, the output is a reverse dense deformation field. ; (2) Local adaptive sensing module: This module is connected to the output of the forward branch and is composed of a multilayer sensing mechanism, with a dense forward deformation field. The data are fed into a multilayer perceptron for feature mapping, which predicts the perception results used to distinguish between truly deformed and undeformed regions. Subsequently, the fine-tuning function module receives the sensing result. With the initial deformation field After fine-tuning, a more accurate high-precision deformation field is output. ; (3) Spatial transformation network: This network includes a grid sampler connected after the fine-tuning function module to receive the fine-tuned high-precision deformation field. and source image As input, the source image is processed using a deformation field mesh. The pixel space is subjected to a differentiable resampling transformation, and the final output is the registered deformed image. ; (4) Multiple loss function constraint mechanism: In the end-to-end connection of the network, multiple loss functions are set to optimize the network weights; specifically, real labels are used at the output of the multilayer perceptron. right Calculate the loss of the perceptual label In fine-tuning deformation field Calculate smoothing loss To ensure the spatial continuity of the deformation field; the registered image is calculated at the network output. With target image loss of similarity between ; and in the forward deformation field With reverse deformation field Apply inverse consistency loss between To improve the consistency and symmetry of deformation estimation under different input orders.

[0065] Step 2 is as follows: Step 2.1: Refer to the printed image Image of the printed matter to be registered As input to the registration deep learning network, the output of the network's forward computation is a two-dimensional dense deformation field mapped from the reference image coordinate system to the coordinate system of the image to be registered, denoted as... For any pixel This represents the predicted two-dimensional displacement vector at that point, where, Indicates in the deformation field Under its influence, specific pixels in the image The deformation displacement vector generated at the location; Represents pixels The displacement generated in the horizontal direction of the two-dimensional coordinate system after network prediction; Represents pixels The displacement generated in the vertical direction of the two-dimensional coordinate system after network prediction; Step 2.2: The network adopts a U-Net encoder-decoder structure, such as... Figure 3As shown, the encoder progressively downsamples to extract multi-scale representations, simultaneously covering local texture details and broader structural semantics; the decoder progressively upsamples to restore spatial resolution and fuses features from the corresponding scales of the encoder through skip connections, thus balancing localization accuracy and deformation continuity when outputting the deformation field; finally, it regresses a dense deformation field with the same resolution as the input. .

[0066] Step 3: Input the reference printed image and the printed image to be registered into the deep learning network to predict the forward deformation field from the reference image to the image to be registered; simultaneously, predict the corresponding reverse deformation field by swapping the image input order, and apply an inverse consistency constraint to the forward and reverse deformation fields to ensure that they satisfy a reciprocal relationship in the sense of spatial mapping, such as... Figure 4 As shown.

[0067] In step 3, the reference printed image will be used. Image of the printed matter to be registered The data is input into a registration deep learning network with shared parameters, and then processed through convolutional feature extraction and decoding regression to obtain the data from... arrive positive deformation field Then swap the input order of the two images, and... Input the inverse uniform deformation network to predict the reverse deformation field ,like Figure 4 As shown, an inverse consistency loss is constructed during training to ensure that the forward and reverse deformation fields satisfy the inverse constraint in the sense of spatial mapping, i.e., by minimizing deformation synthesis. The square norm is made to approach zero, which improves the symmetry and stability of deformation estimation and reduces the registration error caused by changes in input order.

[0068] Step 3 is implemented in the following steps: Step 3.1: Obtain the reference printed image to be registered. Image of the printed matter to be registered The two are then input into a registration deep learning network with shared parameters in a predetermined order, and the output is a positive deformation field. ,in Represents the pixel points in the reference image coordinate system The displacement vector mapped to the coordinate system of the image to be registered; Step 3.2, Exchange and The input order will A registration deep learning network with the same shared parameters outputs an inverse deformation field. ,in Represents the pixel points in the coordinate system of the image to be registered. The displacement vector when mapped back to the reference image coordinate system; Step 3.3: Based on the positive deformation field Calculate pixels Corresponding continuous coordinates in the image to be registered The coordinates u satisfy:

[0069] in The coordinates are non-integer coordinates; Step 3.4: For continuous coordinates Perform neighborhood interpolation sampling to obtain In position Displacement value at; determine The set of integer grid points in the neighborhood For any neighborhood point Calculate the bilinear interpolation weights:

[0070] in and They represent spatial dimensions. The coordinate components on; Step 3.5, define the composite displacement (inverse uniform composite) of the forward and reverse deformation fields as:

[0071] The symbol " " indicates the deformation field synthesis operation; when and When the transformations are inverses, the above synthesis result should approach the zero vector; Step 3.6: Construct the inverse consistency loss function For the domain The composite displacement of all pixels within the area is constrained by the L2 norm to minimize the composite result. The loss function is:

[0072] in It is a norm 2; Step 3.7: During network training, the inverse consistency loss is... As a component of the joint loss function, it participates in backpropagation to simultaneously constrain... and The inverse relationship between the inputs is used to maintain the consistency of the registration results when the input order is changed, and to improve the symmetry and stability of the deformation field estimation.

[0073] Step 4: Input the positive deformation field into the local adaptive sensing module composed of a multilayer perceptron (MLP), determine the pixel-level deformation state, and adaptively adjust the deformation field based on the determination result, such as... Figure 5 As shown, the displacement information of the region determined to be deformed is kept unchanged, while the displacement response of the region determined not to be deformed is weakened, thereby obtaining an optimized deformation field that highlights the real local nonlinear deformation. Step 5: Perform end-to-end optimization of the registration network based on the joint loss function, which includes at least image similarity constraints, deformation field smoothing constraints, inverse consistency constraints, and adaptive perception constraints; and use the optimized deformation field to perform spatial transformation on the printed image to be registered, outputting a registration result aligned with the reference printed image.

[0074] Example 5 This invention relates to an adaptive sensing-based inverse consistency printed image registration method, the flowchart of which is shown below. Figure 1 As shown, please follow these steps: Step 1: Perform local nonlinear deformation enhancement processing on the acquired printed images to construct a printed image registration dataset, and perform uniform size normalization processing on the images to meet the input requirements of the registration network. Part of the dataset is shown in Figure 2. Step 1 is implemented in the following steps: Training and testing datasets were constructed, and local nonlinear data augmentation techniques were applied to synthesize and augment the collected printed image dataset. A printed image registration dataset containing non-rigid deformation features such as local stretching and twisting was constructed. The images were uniformly scaled to 192×160 to meet the network input size requirements. The entire dataset contains 25,200 data pairs, of which the training set contains 20,160 image pairs and the testing set contains 5,040 image pairs. Some results are shown in Figure 2.

[0075] Step 1 is as follows: Step 1.1: Collect images of various types of printed materials, such as medicine boxes, cigarette and alcohol boxes, and food packaging boxes, on a laboratory machine vision platform to form a basic image set. During the acquisition, ensure that the printed content of the reference sample is clear, the exposure and contrast are usable, and record the local deformation phenomena under typical non-ideal working conditions for the design of enhancement strategies. Step 1.2: Employ a local nonlinear deformation enhancement strategy to construct small-area local distortion, stretching, or wrinkling phenomena, and construct a printed image dataset for registration. The reference printed image is denoted as [image name missing]. Generate the image to be registered as follows: : Step 1.2.1, in Randomly crop a candidate local rectangular region. ; Step 1.2.2, for The direction of the inner pixel is random and the amplitude is limited. Random perturbations within a certain range are used to obtain the boundary of the local region after the perturbation, and the perturbation region is constrained not to exceed the boundary; if it exceeds the boundary, it is resampled and generated. Step 1.2.3: Apply the above local disturbance to get This results in the formation of registered sample pairs with local nonlinear deformation. ; Step 1.2.4: Repeat this process. In this way, each reference image can generate multiple images to be registered for different deformation instances, resulting in the final dataset used for training; a portion of the dataset is shown in Figure 2.

[0076] Step 1.3: Size normalization and input normalization. The images are uniformly cropped to the 192×160 size required by the network input and grayscale normalization is performed to meet the tensor size consistency requirements of the network batch training.

[0077] Step 2: Construct a deep learning network for non-rigid registration of printed images, such as... Figure 3 As shown, the network is based on an encoder-decoder structure with shared parameters for predicting dense deformation fields; and inverse consistency deformation constraints and adaptive sensing mechanisms are introduced into the network to enhance the consistency and local discrimination capability of deformation estimation. Step 2 is implemented in the following steps: Construct a deep learning network for print image registration that incorporates an inverse consistency deformation strategy and an adaptive perception mechanism, such as... Figure 3 As shown, the specific structure of a deep learning network is as follows: (1) Shared parameter dual-branch U-Net registration backbone network: This backbone network includes a feature encoder and a feature decoder with skip connections. The network adopts a dual-branch structure, in which the forward branch Source images in ascending order and target image As input, the features are extracted by the encoder through downsampling, and then fused with the decoder through skip connections to output a forward dense deformation field. Simultaneously, the input order of the images is swapped in reverse order. and As input, the reverse branch, which shares network parameters with the forward branch, is fed in. In the middle, the output is a reverse dense deformation field. ; (2) Local adaptive sensing module: This module is connected to the output of the forward branch and is composed of a multilayer sensing mechanism, with a dense forward deformation field. The data are fed into a multilayer perceptron for feature mapping, which predicts the perception results used to distinguish between truly deformed and undeformed regions. Subsequently, the fine-tuning function module receives the sensing result. With the initial deformation field After fine-tuning, a more accurate high-precision deformation field is output. ; (3) Spatial transformation network: This network includes a grid sampler connected after the fine-tuning function module to receive the fine-tuned high-precision deformation field. and source image As input, the source image is processed using a deformation field mesh. The pixel space is subjected to a differentiable resampling transformation, and the final output is the registered deformed image. ; (4) Multiple loss function constraint mechanism: In the end-to-end connection of the network, multiple loss functions are set to optimize the network weights; specifically, real labels are used at the output of the multilayer perceptron. right Calculate the loss of the perceptual label In fine-tuning deformation field Calculate smoothing loss To ensure the spatial continuity of the deformation field; the registered image is calculated at the network output. With target image loss of similarity between ; and in the forward deformation field With reverse deformation field Apply inverse consistency loss between To improve the consistency and symmetry of deformation estimation under different input orders.

[0078] Step 2 is as follows: Step 2.1: Refer to the printed image Image of the printed matter to be registered As input to the registration deep learning network, the output of the network's forward computation is a two-dimensional dense deformation field mapped from the reference image coordinate system to the coordinate system of the image to be registered, denoted as... For any pixel This represents the predicted two-dimensional displacement vector at that point, where, Indicates in the deformation field Under its influence, specific pixels in the image The deformation displacement vector generated at the location; Represents pixels The displacement generated in the horizontal direction of the two-dimensional coordinate system after network prediction; Represents pixels The displacement generated in the vertical direction of the two-dimensional coordinate system after network prediction; Step 2.2: The network adopts a U-Net encoder-decoder structure, such as... Figure 3 As shown, the encoder progressively downsamples to extract multi-scale representations, simultaneously covering local texture details and broader structural semantics; the decoder progressively upsamples to restore spatial resolution and fuses features from the corresponding scales of the encoder through skip connections, thus balancing localization accuracy and deformation continuity when outputting the deformation field; finally, it regresses a dense deformation field with the same resolution as the input. .

[0079] Step 3: Input the reference printed image and the printed image to be registered into the deep learning network to predict the forward deformation field from the reference image to the image to be registered; simultaneously, predict the corresponding reverse deformation field by swapping the image input order, and apply an inverse consistency constraint to the forward and reverse deformation fields to ensure that they satisfy a reciprocal relationship in the sense of spatial mapping, such as... Figure 4 As shown.

[0080] In step 3, the reference printed image will be used. Image of the printed matter to be registered The data is input into a registration deep learning network with shared parameters, and then processed through convolutional feature extraction and decoding regression to obtain the data from... arrive positive deformation field Then swap the input order of the two images, and... Input the inverse uniform deformation network to predict the reverse deformation field ,like Figure 4 As shown, an inverse consistency loss is constructed during training to ensure that the forward and reverse deformation fields satisfy the inverse constraint in the sense of spatial mapping, i.e., by minimizing deformation synthesis. The square norm is made to approach zero, which improves the symmetry and stability of deformation estimation and reduces the registration error caused by changes in input order.

[0081] Step 3 is implemented in the following steps: Step 3.1: Obtain the reference printed image to be registered. Image of the printed matter to be registered The two are then input into a registration deep learning network with shared parameters in a predetermined order, and the output is a positive deformation field. ,in Represents the pixel points in the reference image coordinate system The displacement vector mapped to the coordinate system of the image to be registered; Step 3.2, Exchange and The input order will A registration deep learning network with the same shared parameters outputs an inverse deformation field. ,in Represents the pixel points in the coordinate system of the image to be registered. The displacement vector when mapped back to the reference image coordinate system; Step 3.3: Based on the positive deformation field Calculate pixels Corresponding continuous coordinates in the image to be registered The coordinates u satisfy:

[0082] in The coordinates are non-integer coordinates; Step 3.4: For continuous coordinates Perform neighborhood interpolation sampling to obtain In position Displacement value at; determine The set of integer grid points in the neighborhood For any neighborhood point Calculate the bilinear interpolation weights:

[0083] in and They represent spatial dimensions. The coordinate components on; Step 3.5, define the composite displacement (inverse uniform composite) of the forward and reverse deformation fields as:

[0084] The symbol " " indicates the deformation field synthesis operation; when and When the transformations are inverses, the above synthesis result should approach the zero vector; Step 3.6: Construct the inverse consistency loss function For the domain The composite displacement of all pixels within the area is constrained by the L2 norm to minimize the composite result. The loss function is:

[0085] in It is a norm 2; Step 3.7: During network training, the inverse consistency loss is... As a component of the joint loss function, it participates in backpropagation to simultaneously constrain... and The inverse relationship between the inputs is used to maintain the consistency of the registration results when the input order is changed, and to improve the symmetry and stability of the deformation field estimation.

[0086] Step 4: Input the positive deformation field into the local adaptive sensing module composed of a multilayer perceptron (MLP), determine the pixel-level deformation state, and adaptively adjust the deformation field based on the determination result, such as... Figure 5 As shown, the displacement information of the region determined to be deformed is kept unchanged, while the displacement response of the region determined not to be deformed is weakened, thereby obtaining an optimized deformation field that highlights the real local nonlinear deformation. In step 4, such as Figure 5 As shown, the positive deformation field The input is fed into the adaptive sensing module, where a multilayer perceptron determines the pixel-level deformation state and outputs the deformation region prediction result. ,in accordance with right Adaptive fine-tuning is performed: for regions determined to have undergone deformation, the displacement direction and amplitude information are kept unchanged; for regions determined not to have undergone significant deformation, the displacement intensity is attenuated, weakening its direction and amplitude response, in order to suppress ineffective or redundant deformation and highlight the real local nonlinear deformation regions, thus obtaining the fine-tuned deformation field. .

[0087] Step 4 is implemented in the following steps: Step 4.1: Obtain the positive deformation field output in Step 3. ,in Represents the pixel points in the reference image coordinate system Two-dimensional displacement vector; As input features for the local adaptive sensing module; Step 4.2: The deformation field... Perform pixel-level feature organization, dividing each pixel point The displacement vector component at that location is used as the input feature of that pixel. The data is then input pixel-by-pixel into a multilayer perceptron for deformation state determination; wherein the multilayer perceptron contains at least one hidden layer and a nonlinear activation function, and the linear transformation of its neurons satisfies:

[0088] in As weight, For bias; Step 4.3: Output deformation discrimination results for each pixel using a multilayer perceptron. ,in Indicates whether the pixel belongs to a local deformation region; when When representing "deformation", the pixel point Included in the set of deformation regions ;when When indicating "undeformed", the pixel is... Included in the set of non-deformable regions ; Step 4.4: Deformation judgment results For positive deformation field Perform pixel-level fine-tuning and construct the fine-tuning function. To optimize the deformation field The fine-tuning function satisfies: when the pixel point When a region is identified as a deformation region, the displacement vector remains unchanged; when a pixel... When a region is determined to be non-deformable, the displacement vector is attenuated to weaken invalid deformation. The fine-tuning function expression is as follows:

[0089] in The attenuation coefficient is denoted by ; the preferred value is . .

[0090] Step 4.5: Output the optimized deformation field obtained after fine-tuning. , The original displacement direction and amplitude information are preserved in the deformation region, and the displacement response is suppressed in the non-deformation region, so that the deformation field exhibits more significant local deformation distribution differences in space, thereby improving the ability to capture local nonlinear deformation and reducing the interference of redundant deformation on the registration results. Step 4.6: During the network training phase, deformation label loss is used to constrain the local adaptive perception module: obtain the deformation labels corresponding to the image to be registered. And the deformation labels are matched with the prediction results. Calculate label loss To ensure that the deformation region discrimination result output by the multilayer perceptron is consistent with the deformation label; the label loss is preferably in the form of mean square error.

[0091] in This represents the number of pixels or the number of samples included in the statistics.

[0092] Step 5: Perform end-to-end optimization of the registration network based on the joint loss function, which includes at least image similarity constraints, deformation field smoothing constraints, inverse consistency constraints, and adaptive perception constraints; and use the optimized deformation field to perform spatial transformation on the printed image to be registered, outputting a registration result aligned with the reference printed image.

[0093] Example 6 Objective evaluation metrics were used to test the detection performance of the adaptive perception-based inverse consistency printed image registration method constructed in this invention, in conjunction with Figure 2. Figure 6 The test results are objectively evaluated using the following two indicators: (a) The MSE metric measures the overall difference in pixel intensity between the registered image and the reference image. Let the reference printed image be... The image registered using the method of this invention is Image size is Then MSE is defined as:

[0094] This metric reflects the consistency of the overall grayscale distribution between two images by averaging the squares of the grayscale differences of all pixels. The smaller the MSE value, the lower the pixel-level error between the registered image and the reference image, and the higher the geometric alignment accuracy of the images.

[0095] (b) The SSIM index is an image quality evaluation metric that comprehensively considers brightness, contrast, and structural information, making it more consistent with human visual perception characteristics. Its basic idea is to evaluate registration quality by comparing the similarity of two images at the local structural level. Let the reference image... With registered images The mean values ​​within a certain local window are respectively , The standard deviations are respectively , The covariance is Then SSIM is defined as:

[0096] in, and This is a constant term used to avoid a denominator of zero. The range of values ​​for SSIM is... The closer the value is to 1, the more similar the two images are in terms of structural information.

[0097] Under the same test conditions, the method of this invention was compared with several typical image registration methods for printed materials. The experimental results are shown in Table 1. The present invention achieved the lowest MSE value of 1.88 × 10⁻⁶. -5 It can effectively reduce pixel-level residuals after registration; at the same time, it achieves a higher structural similarity of 0.9964 on the SSIM index, and can better maintain the structural features and visual consistency of printed images in complex local non-rigid deformation scenarios.

[0098] The experimental results above show that the method of the present invention has good performance in both pixel error suppression and structural consistency maintenance, and can effectively meet the registration accuracy requirements of industrial printed images under local non-rigid deformation conditions.

[0099] Table 1 Evaluation results of the registration of this invention on the print registration dataset

[0100] This invention proposes an adaptive perception-based inverse consistency printed image registration method. By constructing a printed image registration dataset incorporating local distortion and stretching perturbations, and introducing a registration network with a shared-parameter encoder-decoder architecture, precise alignment between the reference and target printed images is achieved. During the network's forward pass, the method estimates the forward and reverse dense deformation fields separately, and explicitly models their inverse relationship through inverse consistency constraints, improving the symmetry and stability of deformation estimation. Simultaneously, a pixel-level local adaptive perception mechanism is introduced based on the deformation field to distinguish between truly deformed and non-deformed regions, and to differentially adjust the deformation response, suppressing redundant deformation and highlighting key local deformation areas. Finally, a spatial transformation network is used to achieve geometric correction output for the target printed image, effectively completing pixel-level registration of printed images under local nonlinear deformation conditions, providing a reliable foundation for image spatial information consistency for subsequent defect analysis and quality assessment.

Claims

1. A method for inverse consistency printed image registration based on adaptive perception, characterized in that, The specific steps are as follows: Step 1: Construct a print image registration dataset and normalize the images; Step 2: Construct a deep learning network for non-rigid registration of printed images; Step 3: Predict the positive deformation field from the reference image to the image to be registered; Step 4: Input the positive deformation field into the local adaptive sensing module to obtain an optimized deformation field that highlights the real local nonlinear deformation; Step 5: Perform end-to-end optimization of the registration network based on the joint loss function, and use the optimized deformation field to perform spatial transformation on the printed image to be registered, outputting a registration result aligned with the reference printed image.

2. The method for inverse consistency printed image registration based on adaptive perception according to claim 1, characterized in that, Step 1 is implemented in the following steps: Training and testing datasets were constructed. Local nonlinear data augmentation techniques were applied to synthesize and augment the collected printed image dataset to construct a printed image registration dataset. The images were uniformly scaled to 192×160 to meet the network input size requirements. The entire dataset contains 25,200 data pairs, of which the training set contains 20,160 image pairs and the testing set contains 5,040 image pairs.

3. The method for inverse consistency printed image registration based on adaptive perception according to claim 2, characterized in that, Step 1 is described in detail as follows: Step 1.1: Collect images of various types of printed materials on the laboratory machine vision platform to form a basic image set. During the acquisition, ensure that the printed content of the reference sample is clear, the exposure and contrast are usable, and record the local deformation phenomenon under typical non-ideal working conditions for the design of enhancement strategies. Step 1.2: Employ a local nonlinear deformation enhancement strategy to construct small-area local distortion, stretching, or wrinkling phenomena, and construct a printed image dataset for registration. The reference printed image is denoted as [image name missing]. Generate the image to be registered as follows: : Step 1.2.1, in Randomly crop a candidate local rectangular region. ; Step 1.2.2, for The direction of the inner pixel is random and the amplitude is limited. Random perturbations within a certain range are used to obtain the boundary of the local region after the perturbation, and the perturbation region is constrained not to exceed the boundary; if it exceeds the boundary, it is resampled and generated. Step 1.2.3: Apply the above local disturbance to get This results in the formation of registered sample pairs with local nonlinear deformation. ; Step 1.2.4: Repeat this process. In this way, each reference image can generate multiple images to be registered for different deformation instances, thus obtaining the final dataset used for training. Step 1.3: Size normalization and input normalization. The images are uniformly cropped to the 192×160 size required by the network input and grayscale normalization is performed to meet the tensor size consistency requirements of the network batch training.

4. The method for inverse consistency printed image registration based on adaptive perception according to claim 3, characterized in that, Step 2 is implemented in the following steps: A deep learning network for print image registration, incorporating an inverse consistency deformation strategy and an adaptive perception mechanism, is constructed. The specific structure of the deep learning network is as follows: (1) Shared parameter dual-branch U-Net registration backbone network: This backbone network includes a feature encoder and a feature decoder with skip connections. The network adopts a dual-branch structure, in which the forward branch Source images in ascending order and target image As input, the features are extracted by the encoder through downsampling, and then fused with the decoder through skip connections to output a forward dense deformation field. Simultaneously, the input order of the images is swapped in reverse order. and As input, the reverse branch, which shares network parameters with the forward branch, is fed in. In the middle, the output is a reverse dense deformation field. ; (2) Local adaptive sensing module: This module is connected to the output of the forward branch and is composed of a multilayer sensing mechanism, with a dense forward deformation field. The data are fed into a multilayer perceptron for feature mapping, which predicts the perception results used to distinguish between truly deformed and undeformed regions. ; Subsequently, the fine-tuning function module receives the perception result. With the initial deformation field After fine-tuning, a more accurate high-precision deformation field is output. ; (3) Spatial transformation network: This network includes a grid sampler connected after the fine-tuning function module to receive the fine-tuned high-precision deformation field. and source image As input, the source image is processed using a deformation field mesh. The pixel space is subjected to a differentiable resampling transformation, and the final output is the registered deformed image. ; (4) Multiple loss function constraint mechanism: In the end-to-end connection of the network, multiple loss functions are set to optimize the network weights; specifically, real labels are used at the output of the multilayer perceptron. right Calculate the loss of the perceptual label In fine-tuning the deformation field Calculate smoothing loss To ensure the spatial continuity of the deformation field; the registered image is calculated at the network output. With target image loss of similarity between ; and in the forward deformation field With reverse deformation field Apply inverse consistency loss between To improve the consistency and symmetry of deformation estimation under different input orders.

5. The method for inverse consistency printed image registration based on adaptive perception according to claim 4, characterized in that, Step 2 is described in detail below: Step 2.1: Refer to the printed image Image of the printed matter to be registered As input to the registration deep learning network, the output of the network's forward computation is a two-dimensional dense deformation field mapped from the reference image coordinate system to the coordinate system of the image to be registered, denoted as... For any pixel This represents the predicted two-dimensional displacement vector at that point, where, In the deformation field Under its influence, specific pixels in the image The deformation displacement vector generated at the location; Represents pixels The displacement generated in the horizontal direction of the two-dimensional coordinate system after network prediction; Represents pixels The displacement generated in the vertical direction of the two-dimensional coordinate system after network prediction; Step 2.2: The network adopts a U-Net encoder-decoder structure. The encoder downsamples stepwise to extract multi-scale representations to simultaneously cover local texture details and a wider range of structural semantics. The decoder upsamples stepwise to restore spatial resolution and fuses features from the corresponding scales of the encoder through skip connections, thus balancing localization accuracy and deformation continuity when outputting the deformation field. Finally, a dense deformation field with the same resolution as the input is regressed. .

6. The method for inverse consistency printed image registration based on adaptive perception according to claim 5, characterized in that, In step 3, the reference printed image will be used. Image of the printed matter to be registered The data is input into a registration deep learning network with shared parameters, and then processed through convolutional feature extraction and decoding regression to obtain the data from... arrive positive deformation field Then swap the input order of the two images, and... Input the inverse uniform deformation network to predict the reverse deformation field During training, an inverse consistency loss is constructed to ensure that the forward and reverse deformation fields satisfy inverse constraints in the sense of spatial mapping, i.e., by minimizing deformation synthesis. The square norm is made to approach zero, which improves the symmetry and stability of deformation estimation and reduces the registration error caused by changes in input order.

7. The method for inverse consistency printed image registration based on adaptive perception according to claim 6, characterized in that, Step 3 is implemented in the following steps: Step 3.1: Obtain the reference printed image to be registered. Image of the printed matter to be registered The two are then input into a registration deep learning network with shared parameters in a predetermined order, and the output is a positive deformation field. ,in Represents the pixel points in the reference image coordinate system The displacement vector mapped to the coordinate system of the image to be registered; Step 3.2, Exchange and The input order will A registration deep learning network with the same shared parameters outputs an inverse deformation field. ,in Represents the pixel points in the coordinate system of the image to be registered. The displacement vector when mapped back to the reference image coordinate system; Step 3.3: Based on the positive deformation field Calculate pixels Corresponding continuous coordinates in the image to be registered The coordinates u satisfy: in The coordinates are non-integer coordinates; Step 3.4: For continuous coordinates Perform neighborhood interpolation sampling to obtain In position Displacement value at; determine The set of integer grid points in the neighborhood For any neighborhood point Calculate the bilinear interpolation weights: in and They represent spatial dimensions. The coordinate components on; Step 3.5, define the resultant displacement of the forward and reverse deformation fields as: The symbol " " indicates the deformation field synthesis operation; when and When the transformations are inverses, the above synthesis result should approach the zero vector; Step 3.6: Construct the inverse consistency loss function For the domain The composite displacement of all pixels within the area is constrained by the L2 norm to minimize the composite result. The loss function is: in It is a norm 2; Step 3.7: During network training, the inverse consistency loss is... As a component of the joint loss function, it participates in backpropagation to simultaneously constrain... and The inverse relationship between the inputs is used to maintain the consistency of the registration results when the input order is changed, and to improve the symmetry and stability of the deformation field estimation.

8. The method for inverse consistency printed image registration based on adaptive perception according to claim 7, characterized in that, In step 4, the positive deformation field The input is fed into the adaptive sensing module, where a multilayer perceptron determines the pixel-level deformation state and outputs the deformation region prediction result. ,in accordance with right Adaptive fine-tuning is performed, and for areas determined to have undergone deformation, the displacement direction and amplitude information remain unchanged; For regions determined to have not undergone significant deformation, the displacement intensity is attenuated, weakening its direction and amplitude response to suppress ineffective or redundant deformation and highlight the true local nonlinear deformation region, thus obtaining a fine-tuned deformation field. .

9. The method for inverse consistency printed image registration based on adaptive perception according to claim 8, characterized in that, Step 4 is implemented in the following steps: Step 4.1: Obtain the positive deformation field output in Step 3. ,in Represents the pixel points in the reference image coordinate system Two-dimensional displacement vector; As input features for the local adaptive sensing module; Step 4.2: The deformation field... Perform pixel-level feature organization, dividing each pixel point The displacement vector component at that location is used as the input feature of that pixel. The data is then input pixel-by-pixel into a multilayer perceptron for deformation state determination; wherein the multilayer perceptron contains at least one hidden layer and a nonlinear activation function, and the linear transformation of its neurons satisfies: in As weight, For bias; Step 4.3: Output deformation discrimination results for each pixel using a multilayer perceptron. ,in Indicates whether the pixel belongs to a local deformation region; when When representing "deformation", the pixel point Included in the set of deformation regions ;when When indicating "undeformed", the pixel is... Included in the set of non-deformable regions ; Step 4.4: Deformation judgment results For positive deformation field Perform pixel-level fine-tuning and construct the fine-tuning function. To optimize the deformation field ; The fine-tuning function satisfies: when the pixel point When a region is identified as a deformation region, the displacement vector remains unchanged; when a pixel... When a region is determined to be non-deformable, the displacement vector is attenuated to weaken invalid deformation. The fine-tuning function expression is as follows: in The attenuation coefficient; Step 4.5: Output the optimized deformation field obtained after fine-tuning. , The original displacement direction and amplitude information are preserved in the deformation region, and the displacement response is suppressed in the non-deformation region, so that the deformation field exhibits more significant local deformation distribution differences in space, thereby improving the ability to capture local nonlinear deformation and reducing the interference of redundant deformation on the registration results. Step 4.6: During the network training phase, deformation label loss is used to constrain the local adaptive perception module: obtain the deformation labels corresponding to the image to be registered. And the deformation labels are matched with the prediction results. Calculate label loss To ensure that the deformation region discrimination result output by the multilayer perceptron is consistent with the deformation label; the label loss is preferably in the form of mean square error. in This represents the number of pixels or the number of samples included in the statistics.

10. The method for inverse consistency printed image registration based on adaptive perception according to claim 9, characterized in that, Step 5 is implemented in the following steps: The printed image registration dataset generated by local nonlinear data augmentation is divided into a training dataset and a test dataset, and network input samples are constructed by organizing reference printed images and printed images to be registered in pairs. During the network training phase, the model parameters are iteratively updated based on the principle of stochastic gradient descent to gradually minimize the joint loss function composed of multiple constraints. The joint loss function consists of image similarity loss, deformation field smoothing regularization loss, inverse consistency loss, and deformation region label loss. Each loss term participates in end-to-end backpropagation optimization through weighted combination to collaboratively constrain the learning and generation of the optimal deformation field.

Images