A method for enhancing images of defects in silicon carbide elements based on quantum dot labeling

Abstract

A kind of image enhancement method for defect detection of silicon carbide element based on quantum dot marking.The present application relates to the technical field of semiconductor defect detection and image enhancement, and particularly relates to an image enhancement method for defect detection of silicon carbide element based on quantum dot marking.The present application solves the problems of color distortion, uneven brightness, detail blur and texture loss in the fluorescence image of quantum dot marked silicon carbide element.The method comprises the following steps: generating quantum dot marking on the surface of silicon carbide element using CsPbBr3 precursor solution;collecting digital fluorescence image for defect detection by quantum dot fluorescence imaging system for the surface of silicon carbide element;constructing progressive gated attention fusion enhancement network: global semantic encoding path, local detail extraction path and spatial detail preservation path;the three paths are closely connected through gated feature adapter, and are assisted by progressive supervised attention module to realize stage-by-stage refinement and repair.

Description

Technical Field

[0001] This invention relates to the field of semiconductor defect detection and image enhancement technology, specifically to an image enhancement method for defect detection of silicon carbide components based on quantum dot labeling. Background Technology

[0002] Silicon carbide (SiC), a typical third-generation wide-bandgap semiconductor material, possesses characteristics such as a large bandgap, high breakdown electric field, excellent thermal conductivity, high temperature resistance, and strong radiation resistance, and has been widely used in new energy vehicles, power electronics, rail transportation, and aerospace. However, silicon carbide devices are prone to surface defects such as scratches, pits, and microcracks during processing such as cutting, grinding, and polishing, as well as microscopic defects such as dislocation aggregation and subsurface damage. These defects reduce the quality of the epitaxial layer and the reliability of the device, affecting subsequent manufacturing processes and product yield. Therefore, achieving high-precision detection of surface defects in silicon carbide devices is of great significance for improving processing quality and reducing manufacturing costs.

[0003] Existing defect detection methods for silicon carbide components mainly fall into two categories: destructive testing (DPT) and non-destructive testing (NDT). DPT typically exposes defects through methods such as cutting, etching, and polishing, followed by observation using microscopic equipment. While this method achieves high detection accuracy, it damages the sample and has low efficiency, making it unsuitable for online inspection. NDT primarily utilizes physical signals such as optical, electromagnetic, or ultrasonic signals to detect defects within and on the surface of materials. It has the advantage of not damaging the workpiece, but its detection sensitivity remains limited for small, low-contrast micro- and nano-defects.

[0004] In recent years, quantum dot fluorescent labeling technology has been increasingly applied in the field of micro- and nano-defect detection due to its advantages such as high luminous efficiency, good photostability, tunable fluorescence wavelength, and significant size effect. Among them, CsPbBr3 perovskite quantum dots can form a significant fluorescence enrichment effect in defect regions, creating a high fluorescence contrast between the defect region and the background, providing a new technical approach for the visual detection of minute defects. Compared with traditional fluorescent dyes, CsPbBr3 quantum dots have a higher quantum yield and a narrower emission spectrum, which can effectively improve the sensitivity of defect labeling.

[0005] However, existing quantum dot labeling detection still faces some challenges. On one hand, due to the influence of spin-coating uniformity, crystallization conditions, and surface wettability during the in-situ growth of quantum dots, fluorescence images are prone to uneven brightness distribution, localized overexposure, high background noise, and color shifts. On the other hand, defect edges are susceptible to fluorescence halos caused by localized quantum dot accumulation, resulting in blurred defect outlines and loss of texture details, thus affecting the accuracy of subsequent defect identification. Furthermore, fluorescence microscopy imaging is also affected by factors such as uneven illumination, imaging noise, and optical system issues, leading to color distortion, blurred local details, and loss of high-frequency textures in the acquired defect images.

[0006] Most existing image enhancement methods use single-path convolutional networks or traditional image processing algorithms for enhancement, which makes it difficult to simultaneously achieve overall image color correction, local texture restoration, and high-resolution detail preservation. This often results in problems such as insufficient color correction, incomplete edge restoration, and low utilization of multi-scale features, making it difficult to meet the image quality requirements for subsequent defect detection of quantum dot fluorescent images.

[0007] Therefore, there is an urgent need for an image enhancement method that can fully utilize quantum dot fluorescent labeling information to perform global color correction, local texture restoration, and spatial detail enhancement on fluorescent detection images of silicon carbide components, so as to improve the quality of fluorescent images and provide more accurate and stable image data for subsequent defect detection and identification. Summary of the Invention

[0008] To address the issues of color distortion, uneven brightness, blurred details, and missing textures in the fluorescence images of quantum dot-labeled silicon carbide components, the present invention aims to propose an image enhancement method for defect detection of silicon carbide components based on quantum dot labeling.

[0009] The method includes the following steps: S1. Quantum dot labels are generated on the surface of silicon carbide devices using a CsPbBr3 precursor solution; S2. For the surface of the silicon carbide element after step S1, digital fluorescence images for detecting defects are acquired using a quantum dot fluorescence imaging system. S3. Construct a progressively gated attention fusion enhancement network to enhance digital fluorescence images: The enhanced network includes: a global semantic encoding path, a local detail extraction path, and a spatial detail preservation path; The global semantic coding path receives a digital fluorescence image and outputs four global feature maps with the same spatial size as the digital fluorescence image. , , and ; This is referred to as the first stage of enhanced features. , The digital fluorescence image is input into the first progressively supervised attention module to obtain the first-stage enhanced image. ; Local detail extraction path receives digital fluorescence images and The system then performs double convolution feature extraction and three downsampling encoding processes sequentially to obtain encoded features at three different scales. , and ; Four gated adapter modules are set between the global semantic encoding path and the local detail extraction path to adaptively inject the multi-layer global features output from the global semantic encoding path into the local detail extraction path; the local detail extraction path also includes three cascaded upsampling decoding steps. The input for the first upsampling decoding is used as the input. The result of each upsampling decoding is passed through the corresponding skip connection module and concatenated with the encoded features at the corresponding scale to obtain three decoded features at different scales. , and ; After processing by the first Conv module, the second-stage enhanced features are obtained. , The digital fluorescence image is input along with a second progressively supervised attention module to obtain a second-stage enhanced image. ; Spatial detail preservation path receiving digital fluorescence images and First, pass through the first one After processing by the Conv module and ReLU activation function, the data passes through three cascaded main branches. Each main branch receives the gating fusion results of the corresponding encoded and decoded features from a corresponding gated adapter module. The output of the third main branch is then combined with the digital fluorescence image and... They pass through the convolutional block attention module and the second one in sequence. After processing by the Conv module, the third-stage enhanced image is obtained. .

[0010] Furthermore, the quantum dot fluorescence imaging system includes: an excitation light source module, a microscopic imaging module, and an image acquisition module; The excitation light source module is used to emit excitation light to the silicon carbide element processed in step S1, so that the quantum dot label emits a fluorescent signal. The microscopic imaging module is used to acquire the fluorescence signal emitted by the quantum dots, magnify and image the fluorescence signal, and output a fluorescence image. The image acquisition module is used to receive the fluorescence image output by the microscopic imaging module and convert it into a digital fluorescence image.

[0011] Furthermore, the global semantic encoding path, from input to output, sequentially passes through an image patch segmentation module, a position encoding module, and four cascaded Transformer layers. The sequence features output by each Transformer layer are then processed by the corresponding image reconstruction module to reconstruct a global feature map with the same spatial size as the digital fluorescence image. , , and .

[0012] Furthermore, the first progressive supervised attention module has the same structure as the second progressive supervised attention module; Each progressively supervised attention module fuses the enhanced features of the current stage with the digital fluorescence image. The fused features are then subjected to depthwise separable convolution and normalization in sequence. Based on the normalized features, the attention map of the current stage is generated. The features of the current stage are weighted based on the attention map of the current stage, and residual fusion is performed with the digital fluorescence image to obtain the enhanced image of the current stage.

[0013] Furthermore, each downsampling encoding includes at least one max-pooling downsampling and one double convolutional module.

[0014] Furthermore, each gated feature adapter module receives a global feature map. and the corresponding local feature map of the target Then, channel alignment, spatial alignment, and deep feature extraction are performed sequentially to obtain the current deep features. The current deep features are then subjected to gating and weighting to obtain a weight vector. , This is the index value of the global feature map; Current deep features With weight vector After fusion, the output of the current gated feature adapter module is obtained; in conjunction with the local feature map of the target... Add each element together to obtain the output of the current gated feature adapter module.

[0015] Furthermore, each upsampling decoding includes a bidirectional interpolation module and a dual convolution module; Before the result of each upsampling decoding is connected to the encoded features at the corresponding scale, the encoded features at the corresponding scale are also attention-weighted by the corresponding decoding stage convolutional block attention module to obtain weighted encoded features. The weighted encoded features are then connected to the result of the upsampling decoding.

[0016] Furthermore, for each main branch, parallel residual dense block branches and spatial detail-preserving network convolutional block attention module branches are set up; The outputs of the residual dense block branch and the spatial detail preservation network convolutional block attention module branch are concatenated and then processed by a main branch convolutional block attention module to obtain the output of the current main branch.

[0017] Furthermore, the progressively gated attention fusion enhancement network also includes a loss function. : ,in, For the index value of the stage, For pixel loss, For the first Stage-enhanced images, Real images, For edge loss, For color loss, and These are the hyperparameters corresponding to the loss.

[0018] further, ,in, For L1 loss, for The total number of pixels, ,in, For the Gaussian Laplace operator, It is a constant; ,in, for The channel index value, and They are respectively In the The mean and standard deviation of the channel. and They are respectively In the The mean and standard deviation of the channel.

[0019] The beneficial effects of the method described in this invention are as follows: (1) The present invention uses CsPbBr3 precursor solution to grow in situ on the surface of silicon carbide element, so that quantum dots are preferentially enriched in defect areas such as scratches and pits, and form high-contrast fluorescent markers under fluorescence excitation. Compared with traditional surface detection methods, it can effectively improve the visualization of small defects and provide high-quality raw data for subsequent image enhancement and defect detection.

[0020] (2) The present invention constructs a global semantic coding path, performs long-distance dependency modeling on fluorescence images, obtains global semantic features at different levels, and obtains the first-stage enhanced image through image reconstruction, so that the network can make full use of global color information during the enhancement process, effectively improve the problems of color shift, uneven brightness and inconsistent overall exposure in fluorescence images, and improve the overall imaging quality.

[0021] (3) The present invention uses a gating feature adaptation module to adaptively fuse the global semantic features extracted from the global semantic coding path into the coding features at each level, generates gating weights based on local features, filters the global features and then participates in the fusion, so as to fully preserve the overall semantic information while suppressing the interference of irrelevant features, realize the synergistic enhancement of global color information and local texture information, and improve the recovery ability of defect edges and fine textures.

[0022] (4) The present invention constructs a residual enhancement process through multiple progressive supervised attention modules, so that subsequent stages are based on the enhancement results of the previous stage and gradually compensate for the unrepaired areas, realizing progressive enhancement from coarse to fine, effectively avoiding the omission of details caused by one-time enhancement, and improving the stability of image enhancement.

[0023] (5) The present invention sets up a local detail extraction path and a spatial detail preservation path. The local detail extraction path recovers multi-scale texture information through an encoder-decoder structure, while the spatial detail preservation path further recovers high-frequency details while maintaining the original resolution. It also introduces multi-scale context information by combining cross-stage feature fusion, so that the network can take into account both local texture recovery and spatial detail preservation, effectively reduce the halo effect caused by quantum dot aggregation, and improve the clarity of defect edges.

[0024] (6) The present invention constructs a joint loss function by using pixel loss, edge loss and color loss, and supervises the three enhancement stages simultaneously, so that the enhancement results take into account color consistency, texture continuity and edge clarity, improve the consistency between the enhanced image and the real image, and provide more reliable input for subsequent defect detection. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the quantum dot fluorescence imaging system described in this invention; Figure 2 This is a schematic diagram of the progressive gating attention fusion enhancement network structure described in this invention; Figure 3 This is a schematic diagram of the gated feature adapter structure described in this invention; Figure 4 This is a schematic diagram of the convolutional block attention module structure described in this invention; Figure 5 This is a schematic diagram of the residual dense module structure described in this invention. Detailed Implementation

[0026] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1 This embodiment provides an image enhancement method for defect detection in silicon carbide components based on quantum dot labeling. The method includes the following steps: S1. Quantum dot labels are generated on the surface of silicon carbide devices using a CsPbBr3 precursor solution; The relevant operations in step S1 will be introduced with specific examples: In this embodiment, a quantum dot labeling layer is first generated on the surface of the silicon carbide element to improve the fluorescence response capability of the surface defect area.

[0028] Specifically, the CsPbBr3 precursor solution is prepared using existing mature processes, which can be referred to the method disclosed in Chinese invention patent "An Ion Beam Polishing Method Based on Quantum Dot Coupled Silicon Carbide In-situ Growth" (CN 119890029 A). This embodiment will not elaborate further on this method.

[0029] The prepared CsPbBr3 precursor solution was drop-coated onto the surface of the silicon carbide element to be tested, and the precursor solution was uniformly spread to form a precursor film using a spin-coating process. Subsequently, the silicon carbide element with the precursor film was heated to allow the CsPbBr3 in the precursor solution to nucleate in situ on the surface of the silicon carbide element and grow into a quantum dot labeling layer.

[0030] Because the surface defects of silicon carbide devices, such as scratches, pits, and microcracks, have different microstructures and surface energies than intact surfaces, quantum dots are more likely to accumulate or form differentiated growth in defective areas, resulting in significantly different fluorescence responses between defective and non-defective areas.

[0031] Preferably, after the in-situ growth of quantum dots is completed, the silicon carbide element can be annealed to further promote the stability of the quantum dot crystal structure and improve the luminescence efficiency and stability of the quantum dot labeling layer.

[0032] Finally, a CsPbBr3 quantum dot labeling layer covering the area to be detected is formed on the surface of the silicon carbide element, providing a stable source of fluorescence signal for subsequent fluorescence microscopy imaging.

[0033] S2. For the surface of the silicon carbide element after step S1, digital fluorescence images for detecting defects are acquired using a quantum dot fluorescence imaging system. The relevant operations in step S2 will be introduced with specific examples: like Figure 1 As shown, the quantum dot fluorescence imaging system includes an excitation light source module, a microscopic imaging module, and an image acquisition module. The excitation light source module emits excitation light to the silicon carbide element processed in step S1, causing the quantum dot markers to generate fluorescence signals. Specifically, it includes a light source, a light-collecting lens group, an aperture stop, a condenser lens I, a field stop, a condenser lens II, a rotating four-color fluorescence excitation filter system, an excitation mirror, a dichroic mirror, and an infinity objective lens. The rotating four-color fluorescence excitation filter system includes blue, green, violet, and ultraviolet light filter channels, used to select the appropriate filter channel according to the excitation wavelength of the quantum dots. The microscopic imaging module collects the fluorescence signals emitted by the quantum dots and performs magnified imaging. Specifically, it includes an infinity objective lens, a dichroic mirror, an emitting mirror, and a tube lens. The image acquisition module receives the fluorescence image output by the microscopic imaging module and converts it into a digital fluorescence image. In this embodiment, the image acquisition module uses a CCD camera connected to a computer.

[0034] The specific process of acquiring digital fluorescence images for defect detection using a quantum dot fluorescence imaging system is as follows: The excitation light emitted from the light source passes sequentially through a light-collecting lens group, an aperture stop, condenser lens I, a field stop, and condenser lens II. It then passes through a filter channel in a rotating four-color fluorescence excitation filter system that matches the excitation wavelength of the CsPbBr3 quantum dots to obtain excitation light in the target wavelength range. After its propagation direction is changed by an excitation mirror, the excitation light reaches a dichroic mirror, is reflected by the dichroic mirror, enters the infinity objective lens, and is focused by the infinity objective lens onto the quantum dot labeling layer on the surface of the silicon carbide element to excite the CsPbBr3 quantum dots.

[0035] Upon excitation, the quantum dot labeled layer emits a fluorescence signal of the corresponding wavelength. The fluorescence signal is collected by the infinity objective and returned to the dichroic mirror. After being separated from the excitation light by the dichroic mirror, it is guided to the tube mirror by the emitting mirror. The tube mirror focuses the parallel fluorescent beam output from the infinity objective onto the imaging surface of the CCD camera, forming a fluorescence optical image.

[0036] The CCD camera converts the received fluorescence optical image into a digital fluorescence image and transmits it to a computer. The resulting digital fluorescence image serves as the input image for the subsequent progressively gated attention fusion enhancement network.

[0037] Those skilled in the art can design the composition of the quantum dot fluorescence imaging system according to the characteristics of the actual collected data; this embodiment does not impose any limitations.

[0038] S3. Construct a Progressive Gated Attention Fusion Augmented Network (PGFA-Net) to enhance digital fluorescence images: The relevant operations in step S3 will be introduced with specific examples: like Figure 2 As shown: The enhanced network includes three mutually cooperating feature flow paths: a global semantic encoding path, a local detail extraction path, and a spatial detail preservation path; the three paths are tightly connected through a gated feature adapter (GFA) and supplemented by a progressively supervised attention module (SAM) to achieve stage-by-stage refinement and repair.

[0039] (1) The global semantic coding path receives the digital fluorescence image I and passes through the image block division module, the positional encoding module, and four cascaded Transformer layers in sequence. Each Transformer layer performs global long-distance dependency modeling on the input image, completes overall color shift correction and multi-scale global feature extraction. The sequence features output by each Transformer layer are respectively restored into a global feature map with the same spatial size as the digital fluorescence image by the corresponding image reconstruction head (IRH). , , and Specifically, given an input image ， First, the image block segmentation module divides it into... Non-overlapping image patches, each flattened into a 768-dimensional vector, form a sequence. The learnable location codes are combined with the sequence features output by the image patching module. Element-wise summation is performed to obtain initial sequence features containing spatial location information. These initial sequence features are then sequentially input into four cascaded Transformer layers. Each Transformer layer contains a multi-head self-attention mechanism and a feedforward network to capture long-range dependencies between global pixels. To obtain multi-granularity global semantic information and support subsequent cross-stage feature fusion, the output of the previous Transformer layer is used as the input of the next Transformer layer. To obtain multi-granularity global semantic information and support subsequent cross-stage feature fusion, the sequence features X1, X2, X3, and X4 output from the four Transformer layers are extracted and input into the corresponding image reconstruction heads (IRH).

[0040] Each image reconstruction head consists of a linear mapping layer and a rearrangement layer, which can convert sequence features X i Reconstructed into feature maps with the same spatial dimensions as the input image:

[0041] This is referred to as the first stage of enhanced features. , The digital fluorescence image is input into the first progressively supervised attention module to obtain the first-stage enhanced image. ;and Retain features for subsequent stages of fusion.

[0042] To achieve phased, refined repair, a lightweight progressive supervised attention module (SAM) is designed to control the three-stage enhanced residual update process.

[0043] Each progressively supervised attention module concatenates the current stage's enhanced features with the original digital fluorescence image along the channel dimension. After extracting spatial and channel features through depthwise separable convolution, it uses normalization to obtain intermediate features. These intermediate features are then mapped to a single-channel feature map via a 1×1 convolution and activated by a sigmoid function to generate an attention map. The current stage's enhanced features are weighted according to the attention map and residually fused with the original digital fluorescence image to obtain the enhanced image for the current stage, achieving progressively refined restoration.

[0044] The residual structure forces the model to complete the areas not repaired in the previous stage in a step-by-step manner. SAM is deployed between the first and second stages and between the second and third stages, forming a complete progressive enhancement process.

[0045] (2) The local detail extraction path adopts a U-Net-style encoder-decoder structure and embeds a convolutional block attention module (CBAM) and depth separable convolution to improve the recovery of local texture and edge of silicon carbide element defect detection images.

[0046] The encoding stage of the local detail extraction path uses digital fluorescence image I and first-stage supervised attention features I1. SAM The 6-channel features concatenated along the channel dimension are used as input and first passed through the initial DoubleConv (double convolution) module to obtain the first-scale encoded feature E0, which has a spatial scale of 256×256 and 64 channels. Subsequently, the feature undergoes three downsampling encodings. Each downsampling encoding is processed by max pooling downsampling with a stride of 2 and the DoubleConv module, resulting in second-scale encoding features E1, third-scale encoding features E2, and fourth-scale encoding features E3, with spatial scales of 128×128, 64×64, and 32×32, and channel numbers of 128, 256, and 512, respectively. The input of the first downsampling encoding is obtained by concatenating the first-scale encoding feature E0 with the output of the first GFA module between the global semantic encoding path and the local detail extraction path. In the first downsampling encoding, the output of the max pooling downsampling with a stride of 2 is concatenated with the output of the second GFA module between the global semantic encoding path and the local detail extraction path, and then input into the DoubleConv module. In the second downsampling encoding, the output of the max pooling downsampling with a stride of 2 is concatenated with the output of the third GFA module between the global semantic encoding path and the local detail extraction path, and then input into the DoubleConv module. The output E3 of the third downsampling encoding, concatenated with the output of the fourth GFA module between the global semantic encoding path and the local detail extraction path, serves as the input for the subsequent first upsampling decoding; To enable local feature extraction branches to adaptively utilize global semantic information, a gated feature adapter (GFA) is introduced. Figure 3 As shown, each gated feature adapter module receives a global feature map. and the corresponding target local feature map Then, channel alignment, spatial alignment, and deep feature extraction are performed sequentially to obtain the current deep features (refined features). The current deep features are then subjected to gating and weighting to obtain a weight vector. , This is the index value of the global feature map.

[0047] Current deep features With weight vector After fusion, the output of the current gated feature adapter module is obtained.

[0048] The global semantic encoding path and the local detail extraction path include four GFA modules, where the first GFA module receives the target local feature map. Encode feature E0 for the first scale; The target local feature map received by the second GFA module This is the output of the max-pooling downsampling with a step size of 2 in the first downsampling encoding; The target local feature map received by the third GFA module This is the output of the max-pooling downsampling with a step size of 2 in the second downsampling encoding; The target local feature map received by the fourth GFA module This is the result of the third downsampling encoding.

[0049] Four gated adapter modules are set between the global semantic encoding path and the local detail extraction path, corresponding to the four scale features of the encoding stage of the local detail extraction path, which are used to adaptively inject the multi-layer global features output by the global semantic encoding path into the local detail extraction path.

[0050] The channel alignment: using Convolution will The number of channels is converted to the same as Similarly, the spatial alignment is achieved by using bilinear interpolation. Adjusting the space dimensions to match Consistent. The deep feature extraction: through a Convolution and ReLU activation extract deeper features, resulting in The aforementioned gated weighted processing: for Perform global average pooling, then sequentially pass through two... Convolution (channels are first reduced by a factor of 4 and then restored) and the sigmoid function generate channel weight vectors. The final output is .

[0051] The aforementioned gating mechanism enables the network to automatically filter global information that is beneficial to the learning of current local features, effectively suppressing redundant and disruptive global responses.

[0052] The local detail extraction path also includes three cascaded upsampling decodings. The intermediate results of the first, second, and third upsampling decodings are skip-connected with the corresponding scale-coded features (E3, E2, and E1) weighted by the CBAM module, respectively. The result of each skip connection is then processed by a DoubleConv module to obtain three decoding features D1, D2, and D3 at different scales. The spatial scales of D1, D2, and D3 are restored to 64×64, 128×128, and 256×256, respectively. After D3 is processed by the first Conv module, the second-stage enhanced image I2 is obtained. I2 and the digital fluorescence image I are input together into the second progressive supervised attention module to obtain the second-stage supervised attention feature I2. SAM .

[0053] Each upsampling decoding includes a bidirectional interpolation module. Spatial resolution is gradually restored through bilinear interpolation upsampling to obtain intermediate results for each upsampling decoding. Before the intermediate results of each upsampling decoding are skipped to the corresponding scale of encoded features, the corresponding decoding stage's convolutional block attention module (CBAM) performs attention weighting on the encoded features of the corresponding scale to obtain weighted encoded features. The weighted encoded features are then skipped to the intermediate results of the upsampling decoding.

[0054] The structure of the Convolutional Block Attention Module (CBAM) is as follows: Figure 4 As shown, the convolutional block attention module calculates channel attention and spatial attention sequentially, enabling the network to simultaneously focus on important color channels and key spatial locations.

[0055] (3) To compensate for the loss of high-frequency spatial details caused by downsampling, a Spatial Detail-preserving Network (SDNet) is constructed to achieve full-resolution detail enhancement. SDNet does not include downsampling or upsampling operations and maintains the original 256×256 resolution consistent with the input image throughout the process.

[0056] like Figure 2 As shown, the spatial detail-preserving path receives digital fluorescence images. I Features of second-stage supervised attention I 2 SAMThe features, concatenated along the channel dimension, are sequentially processed through three cascaded backbone branches. Each backbone branch receives the corresponding encoded and decoded features through a gated feature adapter (GFA) module. These encoded and decoded features are then gated and adapted with the input features of the current backbone branch. The gating result is then added element-wise to the input features of the current backbone branch to achieve cross-stage information supplementation. The output of the third backbone branch, after convolutional mapping, is then compared with the digital fluorescence image... I Second-stage supervised attention features I 2 SAM The identity features formed by the concatenation are added element-wise; the resulting fused features are then processed sequentially through a convolutional block attention module and a second Conv module to obtain the third-stage enhanced image. I 3

[0057] like Figure 2 As shown: The output of the Gated Feature Adapter (GFA) module is fused with the current backbone features using an element-wise addition method; Each main branch first adds the output of the previous main branch to the gating adaptation results of the corresponding encoded and decoded features element-wise to obtain the fused input features of the current branch. These fused input features are then input to the Residual Dense Block Branches (RDBs) and the Spatial Details Preserving Network Convolutional Block Attention Module Branches, respectively. The output of the Residual Dense Block Branches is then added element-wise to the output of the Convolutional Block Attention Module Branches, and the result is further refined by the Convolutional Block Attention Module of a main branch to obtain the output of the current main branch.

[0058] Each residual dense block branch stacks 6 residual dense blocks (RDBs). For example... Figure 5 As shown, the input features of each RDB are sequentially processed through six cascaded convolutional units. Each convolutional unit includes a convolutional layer (Conv), a batch normalization layer, and a ReLU activation function. The intermediate features output from each convolutional unit are extracted and concatenated along the channel dimension to obtain densely aggregated features. Subsequently, the following is employed... Convolution performs channel compression and feature fusion on densely aggregated features, and adds the convolution result to the input features element by element to obtain the output features.

[0059] Each spatial detail maintains a branch of the network's convolutional block attention module, including a convolutional block attention module (CBAM). The forward propagation process of SDNet can be represented as: Z0 = ReLU( Conv 3×3 ([I, I2 SAM])) F m =Z m-1+ GFA(E m Z m-1 )+GFA(D m Z m-1 ) Z m = CBAM m (RDB m 6 (F m ) + CBAM′ m (F m m = 1, 2, 3 F out = Conv 3×3 (Z3) + [I, I2 SAM ] I3= Conv 3×3 (CBAM(F out )) in, I This represents the input digital fluorescence image. I 2 SAM Indicates the second stage of image enhancement I 2. Supervised attention features obtained after processing by the second progressive supervised attention module. The spatial detail preservation network first... I and I 2 SAM The data is concatenated along the channel dimension, and then subjected to initial feature extraction and nonlinear mapping via a 3×3 convolution and a ReLU activation function to obtain initial spatial detail features. Z 0. This process can be represented as: Z0 = ReLU( Conv 3×3 ([I, I2 SAM ])) Among them, [I,I2 SAM ] indicates that the input image is concatenated with the second-stage supervised attention features through channels, Conv 3×3 (·) represents a 3×3 convolution operation, and ReLU(·) represents the ReLU activation function. Through this step, the network can simultaneously obtain detail information from the original image and the enhancement information already restored in the second stage before entering the spatial detail preservation branch.

[0060] For the m-th main branch, Z m-1 E represents the input characteristics of this main branch. m and D mThese represent the encoding and decoding stage features corresponding to the main branch, respectively. Since the encoding and decoding features may differ from the current main branch features in terms of channel count and spatial scale, a gated feature adapter (GFA(·)) is first applied to E... m and D m Channel alignment, spatial scale alignment, feature refinement, and gating weighting are performed. The processed encoded and decoded features are then compared with the input features Z of the current backbone branch. m-1 Element-wise addition is performed to obtain the branch input features F after cross-stage gating fusion. m ,Right now: F m =Z m-1+ GFA(E m Z m-1 )+GFA(D m Z m-1 ) Among them, GFA(E m Z m-1 ) indicates that Z m-1 With the scale and number of channels as the target, the encoded feature E m Perform gating adaptation; GFA(D) m Z m-1 ) indicates that Z m-1 With the scale and number of channels as the target, the decoding feature D m Gating adaptation is performed. The purpose of this setting is to enable the spatial detail preservation network to not only rely on its own full-resolution features, but also to make full use of the multi-scale contextual information in the second-stage U-Net encoder and the detail recovery information in the decoder, thereby enhancing the third stage's ability to repair local textures, edge contours and defect details.

[0061] Get F m Subsequently, this feature is entered into the residual dense block branch and the convolutional block attention branch, respectively. Among them, RDB... m 6 (·) represents the residual dense block branch in the m-th main branch. This branch consists of 6 cascaded residual dense modules, used to further extract high-frequency textures and spatial details at full resolution; CBAM′ m (F m The expression () represents the convolutional block attention branch in the m-th main branch, used to apply channel and spatial attention weights to the current fused input features. The outputs of the two branches are element-wise summed and then input into the convolutional block attention module (CBAM) at the end of the main branch. m (·), further filter effective features and suppress redundant responses to obtain the input features Z of the next main branch. m ,Right now: Z m = CBAM m (RDB m 6 (F m ) + CBAM′ m (F m )) Where m=1,2,3 indicates that the spatial detail-preserving network contains 3 cascaded backbone branches. After processing through the three backbone branches, the output feature Z3 of the third backbone branch is obtained.

[0062] Subsequently, the output feature Z3 of the third main branch undergoes channel mapping via a 3×3 convolution, and is then compared with the input image I and the second-stage supervised attention feature I2. SAM The identity features formed by splicing [I,I2] SAM ] Perform element-wise addition to obtain the output fused feature F out ,Right now: F out =Conv 3×3 (Z3)+[I,I2 SAM ] This residual addition structure can preserve the original spatial details of the input and the second-stage recovery results, avoiding excessive smoothing of details or color shift during the third-stage deep feature processing.

[0063] Finally, output the fused feature F. out After undergoing channel and spatial feature recalibration via a convolutional block attention module, and then mapped to a three-channel image via a 3×3 convolution, the third-stage enhanced image I3 is obtained, i.e.: I3=Conv 3×3 (CBAM(F out )) I3 represents the third-stage enhanced image output by the network. Through the above process, the spatial detail preservation network maintains the original spatial resolution without downsampling or upsampling, and further refines defect edges, texture details, and local brightness distribution by combining encoded features, decoded features, and second-stage supervised attention features. To achieve joint optimization of color fidelity, detail clarity, and pixel alignment, three complementary constraints—pixel loss, edge loss, and color loss—are designed, and the output results of the three paths are uniformly supervised.

[0064] (1) L1 pixel loss The loss is used to constrain the pixel-level error between the predicted image and the true sharp image:

[0065] in, for The total number of pixels, For the index value of the stage, For pixel loss, For the first Stage-enhanced images, The images are real-world images, obtained through manual annotation. This loss function ensures the accuracy of the basic reconstruction.

[0066] (2) Edge loss To enhance high-frequency signal preservation, an edge loss based on the Laplacian operator is employed.

[0067] in This represents the Gaussian Laplace operator. Used for numerical stabilization. This loss effectively improves texture and edge sharpness.

[0068] (3) Color loss To address the color shift issue in defect detection images of silicon carbide components, a color loss mechanism based on channel statistical distribution is designed:

[0069] in for The channel index value, and They are respectively In the The mean and standard deviation of the channel. and They are respectively In the The mean and standard deviation of the channels. This loss forces the predicted image to maintain consistency with the color distribution of the real image, effectively correcting the common color cast problem in the detected image.

[0070] (4) Total loss function Output of the three stages , , All are subject to supervision, and the total loss is the weighted sum of the sub-losses at each stage:

[0071] in Hyperparameters are used to balance the losses of each sub-loss.

[0072] In summary, this invention achieves a complete processing flow from acquisition to enhancement of fluorescence images of silicon carbide device defects through the synergistic cooperation of quantum dot in-situ labeling, fluorescence microscopy imaging, and a progressively gated attention fusion enhancement network. Compared with existing technologies, this invention can simultaneously achieve global color correction, local defect texture restoration, and high-resolution spatial detail preservation, effectively improving problems such as color distortion, uneven brightness, blurred edges, and lack of detail in fluorescence images, thus enhancing the quality of enhanced images and providing a more accurate and stable data foundation for subsequent defect detection, localization, and classification.

[0073] Example 2 This embodiment is a further limitation of Embodiment 1.

[0074] To determine the marginal loss weights in the joint loss function With color loss weights To find the optimal combination, a sensitivity experiment was designed using the controlled variable method: one weight was fixed, and the values ​​of the other weight were iterated over, with tests performed on different datasets. To quantitatively analyze the quality of images generated by PGFA, Peak Signal-to-Noise Ratio (PSNR) (Fardo et al. 2016) and Structural Similarity Index (SSIM) were selected as evaluation metrics. PSNR is calculated based on the pixel mean square error; a higher value indicates less image distortion. SSIM measures image similarity from three aspects: brightness, contrast, and structure, which is more consistent with human perception of image structure. Its value ranges from 0 to 1, with values ​​closer to 1 indicating higher image quality.

[0075] The experimental results are shown in Table 1: Table 1 Comparative Analysis of Fixed Weight Combinations

[0076] The first group of experiments had fixed color weights. Gradually adjust the edge loss weights On the dataset, each The peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) showed minimal fluctuations under the given values, indicating stable overall performance. This suggests that appropriate edge constraints can effectively enhance the model's cross-domain generalization ability without sacrificing intra-domain accuracy.

[0077] The second set of experiments had fixed edge loss weights. Gradually adjust color weights On the dataset, The PSNR was highest at 25.29 dB, but the SSIM was lower than that of the PSNR. When approaching; At that time, the lack of color distribution constraints led to incomplete color cast correction; when When the value is increased to 0.10 or 0.15, the intra-domain PSNR improves slightly, but the cross-domain SSIM decreases significantly, and color oversaturation occurs. Considering both intra-domain stability and cross-domain generalization performance, This is the best choice.

[0078] Combining the results of the two sets of experiments, when the edge loss weights Color loss weights When the model achieves optimal overall performance on the dataset, this weight combination is chosen as the parameter setting for the joint loss function. This weight combination effectively balances pixel reconstruction accuracy, edge detail preservation, color fidelity, and model generalization ability.

[0079] All experiments were conducted on Ubuntu 22.04, using the PyTorch 2.8.0 deep learning framework with CUDA version 12.8. The hardware configuration was an NVIDIA RTX 5090 GPU. Input images were uniformly adjusted to 256×256 pixels. The AdamW optimizer was used, with a batch size of 8 and an initial learning rate of 1×10⁻⁶. 4 The training lasted for a total of 100 epochs. No pre-trained weights were used during the training process, and all models were retrained from scratch.

Claims

1. An image enhancement method for defect detection in silicon carbide components based on quantum dot labeling, characterized in that, The method includes the following steps: S1. Quantum dot labels are generated on the surface of silicon carbide devices using a CsPbBr3 precursor solution; S2. For the surface of the silicon carbide element after step S1, digital fluorescence images for detecting defects are acquired using a quantum dot fluorescence imaging system. S3. Construct a progressively gated attention fusion enhancement network to enhance digital fluorescence images: The enhanced network includes: a global semantic encoding path, a local detail extraction path, and a spatial detail preservation path; The global semantic coding path receives a digital fluorescence image and outputs four global feature maps with the same spatial size as the digital fluorescence image. , , and ; This is referred to as the first stage of enhanced features. , The digital fluorescence image is input into the first progressively supervised attention module to obtain the first-stage enhanced image. ; Local detail extraction path receives digital fluorescence images and The system then performs double convolution feature extraction and three downsampling encoding processes sequentially to obtain encoded features at three different scales. , and ; Four gated adapter modules are set between the global semantic encoding path and the local detail extraction path to adaptively inject the multi-layer global features output from the global semantic encoding path into the local detail extraction path; the local detail extraction path also includes three cascaded upsampling decoding steps. The input for the first upsampling decoding is used as the input. The result of each upsampling decoding is passed through the corresponding skip connection module and concatenated with the encoded features at the corresponding scale to obtain three decoded features at different scales. , and ; After processing by the first Conv module, the second-stage enhanced features are obtained. , The digital fluorescence image is input along with a second progressively supervised attention module to obtain a second-stage enhanced image. ; Spatial detail preservation path receiving digital fluorescence images and First, pass through the first one After processing by the Conv module and ReLU activation function, the data passes through three cascaded main branches. Each main branch receives the gating fusion results of the corresponding encoded and decoded features from a corresponding gated adapter module. The output of the third main branch is then combined with the digital fluorescence image and... They pass through the convolutional block attention module and the second one in sequence. After processing by the Conv module, the third-stage enhanced image is obtained. .

2. The image enhancement method for defect detection of silicon carbide components based on quantum dot labeling according to claim 1, characterized in that, The quantum dot fluorescence imaging system includes: an excitation source module, a microscopic imaging module, and an image acquisition module; The excitation light source module is used to emit excitation light to the silicon carbide element processed in step S1, so that the quantum dot label emits a fluorescent signal. The microscopic imaging module is used to acquire the fluorescence signal emitted by the quantum dots, magnify and image the fluorescence signal, and output a fluorescence image. The image acquisition module is used to receive the fluorescence image output by the microscopic imaging module and convert it into a digital fluorescence image.

3. The image enhancement method for defect detection of silicon carbide components based on quantum dot labeling according to claim 2, characterized in that, The global semantic encoding path, from input to output, sequentially passes through an image patch segmentation module, a positional encoding module, and four cascaded Transformer layers. The sequence features output by each Transformer layer are then processed by the corresponding image reconstruction module to reconstruct a global feature map with the same spatial size as the digital fluorescence image. , , and .

4. The image enhancement method for defect detection of silicon carbide components based on quantum dot labeling according to claim 3, characterized in that, The first progressive supervised attention module has the same structure as the second progressive supervised attention module; Each progressively supervised attention module fuses the enhanced features of the current stage with the digital fluorescence image. The fused features are then subjected to depthwise separable convolution and normalization in sequence. Based on the normalized features, the attention map of the current stage is generated. The features of the current stage are weighted based on the attention map of the current stage, and residual fusion is performed with the digital fluorescence image to obtain the enhanced image of the current stage.

5. The image enhancement method for defect detection of silicon carbide components based on quantum dot labeling according to claim 4, characterized in that, Each downsampling encoding includes at least one max-pooling downsampling and one double convolutional module.

6. The image enhancement method for defect detection of silicon carbide components based on quantum dot labeling according to claim 5, characterized in that, Each gated feature adapter module receives a global feature map. and the corresponding local feature map of the target Then, channel alignment, spatial alignment, and deep feature extraction are performed sequentially to obtain the current deep features. The current deep features are then subjected to gating and weighting to obtain a weight vector. , This is the index value of the global feature map; Current deep features With weight vector After fusion, the output of the current gated feature adapter module is obtained; In the local feature map of the target Add each element together to obtain the output of the current gated feature adapter module.

7. The image enhancement method for defect detection of silicon carbide components based on quantum dot labeling according to claim 6, characterized in that, Each upsampling decoding includes a bidirectional interpolation module and a double convolution module; Before the result of each upsampling decoding is connected to the encoded features at the corresponding scale, the encoded features at the corresponding scale are also attention-weighted by the corresponding decoding stage convolutional block attention module to obtain weighted encoded features. The weighted encoded features are then connected to the result of the upsampling decoding.

8. The image enhancement method for defect detection of silicon carbide components based on quantum dot labeling according to claim 7, characterized in that, Each main branch sets up parallel residual dense block branches and spatial detail-preserving network convolutional block attention module branches; The outputs of the residual dense block branch and the spatial detail preservation network convolutional block attention module branch are concatenated and then processed by a main branch convolutional block attention module to obtain the output of the current main branch.

9. The image enhancement method for defect detection of silicon carbide components based on quantum dot labeling according to claim 8, characterized in that, The progressively gated attention fusion enhancement network also includes a loss function. : ,in, For the index value of the stage, For pixel loss, For the first Stage-enhanced images, Real images, For edge loss, For color loss, and These are the hyperparameters corresponding to the loss.

10. The image enhancement method for defect detection of silicon carbide components based on quantum dot labeling according to claim 9, characterized in that, ,in, For L1 loss, for The total number of pixels, ,in, For the Gaussian Laplace operator, It is a constant; ,in, for The channel index value, and They are respectively In the The mean and standard deviation of the channel. and They are respectively In the The mean and standard deviation of the channel.

Images