A deep learning-based fluorescence image denoising method

Abstract

The application discloses a kind of based on deep learning fluorescence image denoising method, comprising the following modules:1.Encoder decoder module, by Enblock, Dnblock stack is formed, and multi-scale fluorescence image features are extracted by layer down sampling, and feature expression is enhanced by built-in LayerNorm, SCA attention, SG activation;2.Multi-scale feature fusion module adopts multi-level feature cross-flow structure, makes deep and shallow layer features bidirectional interaction, and fully fuses different scale space information;3.Gating module, double-branch convolution generates gating weight, and the features of two inputs are weighted and filtered after fusion, adaptively filter noise features, retain effective fluorescence texture information.The technology can not only realize high-quality denoising of multi-scene biological fluorescence image, complete retention of image microscopic details, but also has broad prospects in assisting precise analysis and early diagnosis of disease mechanism.

Description

Technical Field

[0001] This invention belongs to the field of fluorescence image denoising technology, specifically a fluorescence image denoising method based on deep learning. Background Technology

[0002] Fluorescence imaging technology is a core observational tool in life sciences, cell biology, and clinical medicine. With its high specificity and sensitivity, it is widely used in research and clinical settings such as cell structure observation, biomolecular tracking, pathological mechanism research, and early disease screening. However, in actual imaging processes, due to multiple factors including photoelectric noise from imaging equipment, laser scattering, photobleaching, environmental interference, and background interference from the biological sample's own fluorescence, acquired fluorescence images generally suffer from low signal-to-noise ratios, complex noise distributions, and blurred effective details. This noise can mask weak biofluorescence features and destroy image texture information, significantly reducing the imaging quality of fluorescence images and leading to deviations in subsequent image analysis, feature extraction, and pathological interpretation, severely limiting the application of fluorescence imaging technology in high-precision biological research and accurate clinical diagnosis. Therefore, efficient and high-precision fluorescence image denoising technology is crucial for ensuring the authenticity of fluorescence imaging data and improving the accuracy of microscopic biological observation.

[0003] Currently, image denoising techniques are mainly divided into two categories: traditional filtering denoising techniques and deep learning intelligent denoising techniques. Traditional image denoising methods include Gaussian filtering, median filtering, bilateral filtering, and wavelet transform. These algorithms rely on fixed mathematical formulas to smooth image pixels, have simple computational logic, and low implementation costs. They can achieve basic denoising effects for uniform, single Gaussian noise and salt-and-pepper noise. However, traditional algorithms have extremely poor versatility, cannot adapt to the complex mixed noise distribution characteristics of fluorescence images, and suffer from severe detail loss during denoising. They easily weaken the edge features of fluorescence targets, blur microscopic texture details, and exhibit defects such as over-denoising, noise residue, and image distortion, making it difficult to meet the processing requirements of high-precision biological fluorescence imaging.

[0004] With the rapid development of artificial intelligence technology, deep learning-based image denoising methods, with their powerful nonlinear feature fitting and autonomous feature learning capabilities, have broken through the inherent limitations of traditional algorithms and become the mainstream research direction in the field of image denoising. Existing deep learning denoising models are mostly built upon basic convolutional networks and Transformer networks. They learn the differences between effective image features and noise features through massive amounts of data, achieving end-to-end image denoising and reconstruction. Compared to traditional algorithms, their denoising accuracy, scene adaptability, and detail preservation capabilities are all significantly improved.

[0005] However, current mainstream deep learning denoising models still suffer from several unavoidable technical shortcomings when dealing with denoising tasks of biological fluorescence images. Firstly, existing networks mostly employ single-scale feature extraction and fusion methods, only capturing shallow or single-dimensional feature information, failing to fully exploit the multi-scale microscopic features of fluorescence images. Fluorescence images are characterized by tiny target points, fragmented features, and significant differences between shallow and deep features. Single-scale feature extraction easily leads to the loss of crucial details such as tiny fluorescent target points and edge textures. The inability of shallow and deep features to complement each other results in blurred details, insufficient feature representation, and inadequate denoising accuracy in the final denoised image. Secondly, existing deep learning denoising models generally employ fixed and uniform global loss function optimization strategies, without considering the differentiated constraints based on the multi-scale feature distribution characteristics of fluorescence images. Traditional loss functions perform indiscriminate fitting optimization on image target feature regions, edge detail regions, and pure noise background regions, which cannot adapt to the characteristics of fluorescence images: tiny, fragmented target points, large spans between high and low frequency features, and uneven noise distribution. During model training, uniform optimization weights can lead to excessive smoothing of the detailed features of tiny fluorescent targets, weakening the effective texture information after multi-scale fusion. Simultaneously, insufficient noise constraint on complex background regions prevents accurate adaptation to optimization requirements of different scales and signal-to-noise ratios. This ultimately results in low utilization of multi-scale feature fusion, loss of weak fluorescent details, and residual local noise, significantly reducing the overall fidelity and imaging accuracy of denoised fluorescence images. Thirdly, conventional networks lack adaptive feature selection and noise suppression capabilities. Faced with complex mixed noise and densely overlapping fluorescent target scenes in fluorescence images, they cannot accurately distinguish between noise features and effective biological features, easily leading to incomplete denoising, image artifacts, and feature distortion. This makes it difficult to simultaneously meet the dual requirements of efficient denoising and preservation of fine details, greatly limiting the application of fluorescence images in high-precision scientific research and clinical diagnosis.

[0006] To address the shortcomings of existing technologies, this invention proposes a fluorescence image denoising method based on deep learning feature fusion. By constructing a network architecture that combines an encoder-decoder structure, a multi-scale feature fusion module, and a gated adaptive filtering module, this method specifically solves the technical problems of traditional models, such as detail loss, inaccurate optimization, and weak noise discrimination ability. Specific improvements are as follows: First, this invention designs a multi-level encoder module to replace the traditional single-scale feature extraction structure. By extracting shallow texture details and deep semantic features of fluorescence images layer by layer through multiple encoding units, it simultaneously captures fine-grained information of tiny fluorescent target points and global image distribution features. This solves the problems of blurred small target features and loss of microscopic details caused by single downsampling in traditional networks, effectively improving the network's ability to extract multi-scale fluorescence features and providing a sufficient and complete feature foundation for subsequent high-precision denoising.

[0007] Second, this invention adds a progressive multi-scale feature fusion module to perform cross-layer bidirectional interactive fusion of multi-level features of different scales output by the encoder. Through the stepwise association and complementary update of shallow, medium and deep features, the full flow and aggregation of high and low frequency features are realized, avoiding the problem of one-sided feature information at a single level, maximizing the utilization of effective multi-scale features, making up for the shortcomings of insufficient feature fusion and poor scale adaptability of traditional models, and significantly improving the utilization rate of multi-scale features.

[0008] Third, this invention introduces a gated adaptive module to achieve intelligent discrimination and weighted calibration of effective features and noise features in fluorescence images. The gated module can adaptively generate weight coefficients based on the feature differences of different image regions, enhancing and preserving effective features such as fluorescence targets and edge details, while suppressing and filtering complex background noise and redundant pseudo-features. This solves the problems of incomplete noise reduction, feature distortion, and severe artifacts caused by the indiscriminate feature processing of traditional networks. Through multi-module collaborative optimization, this invention can achieve both efficient noise removal and high-precision preservation of microscopic details in complex fluorescence imaging scenarios such as high-intensity noise, dense target overlap, and low signal-to-noise ratio, significantly improving the denoising quality and imaging fidelity of fluorescence images. Summary of the Invention

[0009] In view of the above-mentioned technical problems existing in the prior art, the present invention discloses a fluorescence image denoising method based on deep learning.

[0010] The present invention adopts the following technical solution: A deep learning-based fluorescence image denoising method includes the following modules: The encoder-decoder module consists of stacked Enblock and Dnblock components. It extracts multi-scale fluorescence image features through layer-by-layer downsampling and incorporates LayerNorm, SCA attention, and SG activation to enhance feature representation. Multi-scale feature fusion module: It adopts a multi-level feature cross-flow structure, which allows deep and shallow features to interact bidirectionally and fully integrate spatial information at different scales; Gating module: Dual-branch convolution generates gating weights, which are then weighted, filtered, and fused to adaptively filter noisy features and retain effective fluorescent texture information.

[0011] Preferably, the encoder-decoder module serves as the foundational feature extraction and image reconstruction unit for the entire denoising network. The encoder comprises multi-level stacked coding units, which progressively downsample and layer-wise mine multi-dimensional feature information from the fluorescence image. Shallow coding units focus on learning detailed features such as image edges, textures, and tiny fluorescent targets, while deep coding units focus on extracting global semantic and background distribution features, effectively addressing the problems of detail loss and insufficient feature representation caused by single-scale feature extraction in traditional networks.

[0012] The specific steps in the encoder-decoder module are as follows: The input raw fluorescence data is passed through a 3×3 convolutional layer for preliminary feature extraction, and the number of channels is increased from the original 3 to 16.

[0013] Then, after a 1×1 ordinary convolution, the number of channels increases from 16 to 32, and the information of each channel is associated. Then, after a 3×3 depthwise convolution with a stride of 2, it is downsampled.

[0014] Then, the input data passes through the Enblock module, whose specific steps are as follows: 1. Layernorm is applied to the input data to make the model training more stable, and then the number of channels is doubled through a 1×1 ordinary convolution, followed by a 3×3 depthwise convolution; then, the SG module separates the input data into two feature maps with equal number of feature channels, and then performs a dot product operation; then, the SCA module simplifies the channel attention mechanism, first performing global pooling on the input data, then inputting a 1×1 ordinary convolution, and finally multiplying it with the original input data; then, a 1×1 ordinary convolution is applied, and a residual connection is made with the input data to obtain the intermediate output feature result; then, Layernorm, a 1×1 ordinary convolution, the SG module, a 1×1 ordinary convolution, and a residual connection is made with the obtained result and the intermediate output feature result.

[0015] Then, after three combinations of downsampling and Enblock modules, and finally one Enblock module, the encoder part is completed.

[0016] Then, the data passes through a Deblock module. The Deblock module's structure is as follows: the input data is processed through a Layernorm, followed by 1×1 regular convolutions, 3×3 depthwise convolutions, 5×5 depthwise convolutions, and 7×7 depthwise convolutions. The outputs of these three convolutions at different scales are weighted and fused. Then, it passes through a 1×1 regular convolution, an SG module, an SCA module, and another 1×1 regular convolution. The output is then residually concatenated with the original input to obtain intermediate output features. Next, the Layernorm module processes the data through 1×1 regular convolutions, 1×1 regular convolutions, 1×1 regular convolutions, and 3×3 depthwise convolutions. The resulting two feature maps are multiplied, then processed through a 1×1 regular convolution, and finally residually concatenated with the intermediate output features. Finally, the data undergoes upsampling and cross-connection. The Deblock module is combined four times, with the number of feature channels halved each time.

[0017] Finally, a 3×3 ordinary convolution is performed to obtain the output.

[0018] Preferably, in the multi-scale feature fusion module, a progressive fusion approach can achieve full flow and deep coupling of shallow detailed features, mid-level transitional features, and deep global semantic features of fluorescence images. This allows features at each scale to complement each other and be corrected layer by layer, effectively solving the problems of fragmented shallow and deep features, single utilization of scale information, and insufficient fusion of small fluorescent target features in traditional denoising networks. Through multi-level cross-iterative fusion, fine details such as weak fluorescent targets and edge textures at low scales are preserved, while high-scale global semantic features are fully utilized to suppress complex background noise. This significantly improves the utilization rate of effective features at multiple scales, providing rich and complete feature representations for subsequent gating screening and high-precision image reconstruction, and significantly enhancing the denoising and detail restoration capabilities of complex low signal-to-noise ratio fluorescence images.

[0019] The specific steps in the multi-scale feature fusion module are as follows: Multi-scale features X1, X2, X3, and X4 from each level of the encoder's Enblock output are obtained, and feature fusion is performed using a progressive cross-layer interaction strategy. Shallow feature X1 and sub-shallow feature X2 are bidirectionally fused to obtain X1 and X2; After fusion and updating, X1 and X2 are obtained by fusion of X1 and X2, X2 is obtained by cross-fusion of X2 with X1 and X3, and X3 is obtained by fusion of X3 with X2. After fusion and updating, X1, X2, and X3 are obtained. The fusion of X1 and X2 is the same as in step 3. X3 is cross-fused with X2 and X4 to obtain X3. X4 is fused with X3 and X4 to obtain X4. By enabling cross-scale information flow layer by layer, features of deep and shallow layers are fully complementary, thus preserving the texture and molecular details of fluorescence images at different scales.

[0020] Preferably, in the gating module, the gating module can accurately distinguish between effective target areas, edge detail areas, and pure noise background areas in the fluorescence image. It enhances the weight of weak fluorescence textures and small target features carrying biological information, thereby strengthening the characterization capability of key effective features. At the same time, it suppresses and removes cluttered background noise, fused redundant pseudo-features, and invalid high-frequency interference information. This adaptive gating control strategy can dynamically adapt to complex fluorescence imaging scenarios with different noise intensities, target density, and signal-to-noise ratios.

[0021] The specific steps in the gating module are as follows: For the output of the feature fusion part, gated fusion is performed with the corresponding part of the decoder. The output of the feature fusion module is concatenated with the output of the corresponding decoder in the feature channel direction. Then, the results of the two branches are multiplied by passing a 1×1 ordinary convolution, a 3×3 depthwise convolution and GELU activation respectively. Then, it is passed through a 1×1 ordinary convolution and residually connected with the data of the input module.

[0022] The model training steps are as follows: The network input is the original three-channel fluorescence image. First, it passes through a 3×3 standard convolutional layer. The convolution operation keeps the image size unchanged and maps the three feature channels of the original input image into 16-dimensional feature channels, completing the initial shallow feature extraction and channel dimensionality upscaling.

[0023] The 16-dimensional features obtained from the preprocessing are sequentially passed through a combination module of 1×1 standard convolution and 3×3 depth convolution to double the number of feature channels and halve the size of the feature map space, thus completing the first downsampling and feature dimension expansion.

[0024] The above output features are fed into the Enblock encoding layer to obtain the first-level shallow feature X1; then, through a combination module of 1×1 standard convolution and 3×3 depth convolution, the feature channels are doubled and the feature map size is halved to achieve the second downsampling operation.

[0025] The features after the second downsampling are input into the Enblock encoding layer, and the second-level mid-layer features X2 are output. The combination structure of 1×1 standard convolution and 3×3 depth convolution is used to complete the third downsampling process, which doubles the feature channel capacity and halves the feature map size.

[0026] The processed features are input into the Enblock encoding layer to obtain the third-level mid-deep features X3; then, through a combination module of 1×1 standard convolution and 3×3 depth convolution, the fourth downsampling operation is completed, doubling the feature channels and halving the size.

[0027] The fourth downsampled feature is input into the Enblock encoding layer, which outputs the fourth level deep feature X4. Finally, it undergoes a combination of 1×1 standard convolution and 3×3 depth convolution for downsampling, which further compresses the feature size and expands the feature channels.

[0028] The deep features obtained by the final downsampling are sequentially passed through the Enblock encoding unit and the Deblock denoising and enhancement unit to complete the semantic enhancement, noise filtering and feature purification of the deep features, which constitute the feature output of the network bottleneck layer.

[0029] Upsampling is performed on the bottleneck layer features, and channel compression is performed in conjunction with 1×1 standard convolution to halve the number of feature channels and double the feature map size, thus completing the first decoding dimension restoration.

[0030] The X4 features optimized by the multi-scale feature fusion module and the original encoded X4 features are input into the gating module. After gating adaptive weight calibration, the features are added point by point to complete the cross-layer skip connection. Finally, the feature is connected to the Deblock layer to complete feature repair.

[0031] The refined features are upsampled again and subjected to 1×1 standard convolution operations to halve the feature channels and double the size. The X3 features optimized by multi-scale fusion are input into the gating module for calibration and fusion with the original encoded X3 features. Feature superposition is completed through skip connections, and finally the output is optimized by the Deblock layer.

[0032] Continue performing upsampling and 1×1 standard convolution operations to halve the number of channels and double the image size; after the multi-scale fusion optimized X2 features and the original encoded X2 features are adaptively filtered and calibrated by the gating module, feature fusion is completed by skip connections, and then the features are refined by the Deblock layer.

[0033] The final upsampling and 1×1 standard convolution are performed to restore the original spatial size of the image. The multi-scale fused and optimized X1 features and the original shallow X1 features are input into the gating module to complete the precise calibration. The shallow details are supplemented by skip connections, and finally the full-scale feature optimization and reconstruction is completed by the Deblock layer.

[0034] The final feature map after decoding and reconstruction is fed into a 3×3 standard convolutional layer to complete feature channel mapping and final image detail restoration, and outputs the final denoised high-definition fluorescence image.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention uses a symmetrical encoding and decoding structure to extract multi-scale shallow and deep features of fluorescence images in layers, effectively preserving tiny target points and texture details, thereby improving the feature extraction and detail reconstruction capabilities of low signal-to-noise ratio fluorescence images.

[0036] 2. This invention utilizes a progressive multi-scale feature fusion structure to bidirectionally fuse features at all levels, compensate for the defects of feature fragmentation between shallow and deep areas, and fully aggregate effective information from multiple scales, thereby significantly improving the image denoising effect in complex backgrounds and dense target scenes.

[0037] 3. This invention introduces a gated adaptive module to adaptively identify and enhance effective fluorescence features, suppress background noise and redundant false features, thereby effectively improving the denoising accuracy and imaging fidelity of fluorescence images. Attached Figure Description

[0038] Figure 1 This is a diagram of the overall network structure in this invention. Figure 2 : This is a structural diagram of the Enblock in this invention; Figure 3 : This is a structural diagram of the Deblock in this invention; Figure 4 : This is a structural diagram of the MFF in this invention; Figure 5This is a structural diagram of the GCF in this invention. Detailed Implementation

[0039] To more clearly illustrate the technical solution of the present invention, the following detailed description is provided in conjunction with specific embodiments.

[0040] A preferred embodiment of the present invention proposes a deep learning-based fluorescence image denoising method, the algorithm structure of which is as follows: Figure 1 As shown.

[0041] Combination Figure 1-5 The specific implementation steps of the deep learning-based fluorescence image denoising method in this embodiment are as follows: The network input is the original three-channel fluorescence image. First, it passes through a 3×3 standard convolutional layer. The convolution operation keeps the image size unchanged and maps the three feature channels of the original input image into 16-dimensional feature channels, completing the initial shallow feature extraction and channel dimensionality upscaling.

[0042] The 16-dimensional features obtained from the preprocessing are sequentially passed through a combination module of 1×1 standard convolution and 3×3 depth convolution to double the number of feature channels and halve the size of the feature map space, thus completing the first downsampling and feature dimension expansion.

[0043] The above output features are fed into the Enblock encoding layer to obtain the first-level shallow feature X1; then, through a combination module of 1×1 standard convolution and 3×3 depth convolution, the feature channels are doubled and the feature map size is halved to achieve the second downsampling operation.

[0044] The features after the second downsampling are input into the Enblock encoding layer, and the second-level mid-layer features X2 are output. The combination structure of 1×1 standard convolution and 3×3 depth convolution is used to complete the third downsampling process, which doubles the feature channel capacity and halves the feature map size.

[0045] The processed features are input into the Enblock encoding layer to obtain the third-level mid-deep features X3; then, through a combination module of 1×1 standard convolution and 3×3 depth convolution, the fourth downsampling operation is completed, doubling the feature channels and halving the size.

[0046] The fourth downsampled feature is input into the Enblock encoding layer, which outputs the fourth level deep feature X4. Finally, it undergoes a combination of 1×1 standard convolution and 3×3 depth convolution for downsampling, which further compresses the feature size and expands the feature channels.

[0047] The deep features obtained by the final downsampling are sequentially passed through the Enblock encoding unit and the Deblock denoising and enhancement unit to complete the semantic enhancement, noise filtering and feature purification of the deep features, which constitute the feature output of the network bottleneck layer.

[0048] Upsampling is performed on the bottleneck layer features, and channel compression is performed in conjunction with 1×1 standard convolution to halve the number of feature channels and double the feature map size, thus completing the first decoding dimension restoration.

[0049] The X4 features optimized by the multi-scale feature fusion module and the original encoded X4 features are input into the gating module. After gating adaptive weight calibration, the features are added point by point to complete the cross-layer skip connection. Finally, the feature is connected to the Deblock layer to complete feature repair.

[0050] The refined features are upsampled again and subjected to 1×1 standard convolution operations to halve the feature channels and double the size. The X3 features optimized by multi-scale fusion are input into the gating module for calibration and fusion with the original encoded X3 features. Feature superposition is completed through skip connections, and finally the output is optimized by the Deblock layer.

[0051] Continue performing upsampling and 1×1 standard convolution operations to halve the number of channels and double the image size; after the multi-scale fusion optimized X2 features and the original encoded X2 features are adaptively filtered and calibrated by the gating module, feature fusion is completed by skip connections, and then the features are refined by the Deblock layer.

[0052] The final upsampling and 1×1 standard convolution are performed to restore the original spatial size of the image. The multi-scale fused and optimized X1 features and the original shallow X1 features are input into the gating module to complete the precise calibration. The shallow details are supplemented by skip connections, and finally the full-scale feature optimization and reconstruction is completed by the Deblock layer.

[0053] The final feature map after decoding and reconstruction is fed into a 3×3 standard convolutional layer to complete feature channel mapping and final image detail restoration, and outputs the final denoised high-definition fluorescence image. It should be noted that the aforementioned implementation examples are merely typical demonstrations illustrating the core principles of this invention, and are not intended to limit the technical boundaries of this invention. While maintaining the integrity of the core ideas of this invention, those skilled in the art can make adaptive improvements or innovative extensions based on specific application scenarios, and such derivative solutions should still be considered an integral part of the scope of patent protection of this invention.

Claims

1. A deep learning-based fluorescence image denoising method, characterized by: Includes the following modules: The encoder-decoder module consists of stacked Enblock and Dnblock components. It extracts multi-scale fluorescence image features through layer-by-layer downsampling and incorporates LayerNorm, SCA attention, and SG activation to enhance feature representation. Multi-scale feature fusion module: It adopts a multi-level feature cross-flow structure, which allows deep and shallow features to interact bidirectionally and fully integrate spatial information at different scales; Gating module: Dual-branch convolution generates gating weights, which are then weighted, filtered, and fused to adaptively filter noisy features and retain effective fluorescent texture information.

2. The deep learning-based fluorescence image denoising method as described in claim 1, characterized in that: The specific steps in the encoder-decoder module are as follows: The input raw fluorescence data is passed through a 3×3 convolutional layer for preliminary feature extraction, and the number of channels is increased from the original 3 to 16. Then, after a 1×1 ordinary convolution, the number of channels increases from 16 to 32, and the information of each channel is associated. Then, after a 3×3 depthwise convolution with a stride of 2, it is downsampled. Then, the input data passes through the Enblock module, whose specific steps are as follows:

1. Layernorm is applied to the input data to make the model training more stable, and then the number of channels is doubled through a 1×1 ordinary convolution, followed by a 3×3 depthwise convolution; then, the input data passes through the SG module, which separates the input data into two feature maps with equal number of feature channels, and then performs a dot product operation; then, the input data passes through the SCA module, which simplifies the channel attention mechanism. First, global pooling is performed on the input data, then a 1×1 ordinary convolution is applied, and finally, it is multiplied with the original input data; then, a 1×1 ordinary convolution is applied, and residual connections are made with the input data to obtain the intermediate output feature results; Then, after passing through Layernorm, a 1×1 ordinary convolution, the SG module, and a 1×1 ordinary convolution, the result is residually connected with the intermediate output feature result. Then, after three combinations of downsampling and Enblock modules, and finally one Enblock module, the encoder part is completed. Then, the data passes through a Deblock module. The Deblock module's structure is as follows: the input data is processed through a Layernorm, followed by 1×1 regular convolutions, 3×3 depthwise convolutions, 5×5 depthwise convolutions, and 7×7 depthwise convolutions. The outputs of these three convolutions at different scales are weighted and fused. Then, it passes through a 1×1 regular convolution, an SG module, an SCA module, and another 1×1 regular convolution. The output is then residually concatenated with the original input to obtain intermediate output features. Next, the Layernorm module processes the data through 1×1 regular convolutions, 1×1 regular convolutions, 1×1 regular convolutions, and 3×3 depthwise convolutions. The resulting two feature maps are multiplied, then processed through a 1×1 regular convolution, and finally residually concatenated with the intermediate output features. Finally, the data undergoes upsampling and cross-connection. The Deblock module is combined four times, with the number of feature channels halved each time. Finally, a 3×3 ordinary convolution is performed to obtain the output.

3. The deep learning-based fluorescence image denoising method as described in claim 1, characterized in that: Multi-scale feature fusion module. The specific steps are as follows: Multi-scale features X1, X2, X3, and X4 from each level of the encoder's Enblock output are obtained, and feature fusion is performed using a progressive cross-layer interaction strategy. Shallow feature X1 and sub-shallow feature X2 are bidirectionally fused to obtain X1 and X2; After fusion and updating, X1 and X2 are obtained by fusion of X1 and X2, X2 is obtained by cross-fusion of X2 with X1 and X3, and X3 is obtained by fusion of X3 with X2. After fusion and updating, X1, X2, and X3 are obtained. The fusion of X1 and X2 is the same as in step 3. X3 is cross-fused with X2 and X4 to obtain X3. X4 is fused with X3 and X4 to obtain X4. By enabling cross-scale information flow layer by layer, features at both shallow and deep layers are fully complemented, thus preserving the texture and molecular details of fluorescence images at different scales.

4. The deep learning-based fluorescence image denoising method as described in claim 4, characterized in that: The gating module, the specific steps are as follows: For the output of the feature fusion part, gated fusion is performed with the corresponding part of the decoder. The output of the feature fusion module is concatenated with the output of the corresponding decoder in the feature channel direction. Then, the results of the two branches are multiplied by passing a 1×1 ordinary convolution, a 3×3 depthwise convolution and GELU activation respectively. Then, it is passed through a 1×1 ordinary convolution and residually connected with the data of the input module.

Images