An eye fundus image optimization method, a model deployment method, and an eye fundus imaging system

Abstract

This application discloses a fundus image optimization method, a model deployment method, and a fundus imaging system, including: acquiring an original fundus image to be optimized; optimizing multiple frequency domain feature components obtained after decomposing the original fundus image, and obtaining a reconstructed feature image based on the optimized multiple frequency domain feature components; performing hierarchical upsampling on the reconstructed feature image, and obtaining a high-resolution image through channel grouping, intra-group attention mechanism, and cross-group feature fusion; and performing collaborative enhancement of the high-resolution feature image based on spatial attention mechanism and channel attention mechanism to obtain an optimized target fundus image. This application improves image contrast and clarity; accurately restores rich structural and texture details while increasing image resolution; adaptively highlights key features with diagnostic value and further suppresses background interference, ultimately generating a high-quality target fundus image.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to an eye fundus image optimization method, a model deployment method and an eye fundus imaging system. BACKGROUND

[0002] The eye is an important window reflecting the health condition of the human body. The eye fundus is the only part of the human body where blood vessels and nerves can be directly observed non-invasively. The clarity of the eye fundus image is directly related to the accuracy of early screening and diagnosis of various ophthalmic diseases (such as diabetic retinopathy, glaucoma, macular degeneration, etc.) and even systemic diseases. Therefore, obtaining high-quality eye fundus images is a key link in clinical ophthalmic examination. However, the eye fundus images collected in the related art have problems such as blurred details and insufficient contrast, which affect the accuracy of clinical diagnosis. Therefore, how to improve the quality of the eye fundus image is a problem to be solved at present. SUMMARY

[0003] The embodiments of the present application provide an eye fundus image optimization method, a model deployment method and an eye fundus imaging system, which solve the technical problem that the eye fundus images collected in the prior art have blurred details and insufficient contrast, resulting in low quality of the eye fundus images, and achieve the technical effect of enhancing the details and contrast of the eye fundus image and improving the quality of the eye fundus image.

[0004] In a first aspect, the present application provides an eye fundus image optimization method, which comprises: obtaining an original eye fundus image to be optimized; optimizing a plurality of frequency domain feature components obtained by decomposing the original eye fundus image, and obtaining a reconstructed feature image based on the plurality of optimized frequency domain feature components; performing hierarchical upsampling on the reconstructed feature image, and obtaining a high-resolution image through channel grouping, in-group attention mechanism and cross-group feature fusion; performing collaborative enhancement on the high-resolution feature image based on a spatial attention mechanism and a channel attention mechanism, and obtaining an optimized target eye fundus image.

[0005] Further, the step of optimizing a plurality of frequency domain feature components obtained by decomposing the original eye fundus image, and obtaining a reconstructed feature image based on the plurality of optimized frequency domain feature components comprises: decomposing the original eye fundus image to obtain a plurality of frequency domain feature components; performing depth separable convolution processing on each frequency domain feature component to obtain each optimized frequency domain feature component; reconstructing the plurality of optimized frequency domain feature components to obtain the reconstructed feature image.

[0006] Furthermore, the multiple frequency domain feature components obtained after decomposing the original fundus image are optimized respectively, and a reconstructed feature image is obtained based on the optimized multiple frequency domain feature components, including: The original fundus image is decomposed using wavelet transform to obtain multiple frequency domain feature components; the multiple frequency domain feature components include at least low-frequency approximation components, horizontal detail components, vertical detail components, and diagonal detail components; Each frequency domain feature component is subjected to depthwise separable convolution to obtain the optimized frequency domain feature components. The reconstructed feature image is obtained by reconstructing multiple frequency domain feature components using inverse wavelet transform.

[0007] Furthermore, the reconstructed feature image is subjected to hierarchical upsampling, and a high-resolution image is obtained through channel grouping, intra-group attention mechanism, and cross-group feature fusion, including: Multiple cascaded upsampling blocks are used to perform hierarchical upsampling on the reconstructed feature image based on depth-separable convolution and pixel shuffling to obtain an initial high-resolution image; The initial high-resolution image is grouped into channels according to multiple preset channel features to obtain feature sub-images corresponding to each channel feature. Based on the response intensity of each channel in the current feature map, determine the weight coefficient of each channel in each feature map; Based on the weight coefficients of each feature sub-map, the corresponding feature sub-map is recalibrated to obtain the enhanced feature sub-map; Cross-group feature fusion is performed on each enhanced feature sub-image to obtain the high-resolution image.

[0008] Furthermore, the high-resolution feature image is collaboratively enhanced based on spatial attention and channel attention mechanisms to obtain an optimized target fundus image, including: Based on the spatial attention mechanism, the spatial attention weight map corresponding to the high-resolution feature image is determined; Based on the channel attention mechanism, the channel attention weight map corresponding to the high-resolution feature image is determined; The spatial attention weight map and the channel attention weight map are fused to obtain the fused attention weight; The high-resolution feature image is calibrated based on the fused attention weights to obtain the target fundus image.

[0009] Furthermore, based on the spatial attention mechanism, the spatial attention weight map corresponding to the high-resolution feature image is determined, including: Max pooling along the channel dimension is performed on the high-resolution feature image to obtain a salient feature map; The high-resolution feature image is subjected to average pooling along the channel dimension to obtain an average feature map; The salient feature map and the average feature map are concatenated along the channel dimension to obtain a fused feature map; The fused feature map is sequentially input into a standard convolutional layer and a Sigmoid activation function to obtain a spatial attention weight map.

[0010] Furthermore, based on the channel attention mechanism, the channel attention weight map corresponding to the high-resolution feature image is determined, including: Global average pooling is performed on the high-resolution feature image to obtain the channel statistics vector; The channel statistical vectors are input into the bottleneck structure neural network to obtain the intermediate vectors representing the dependencies between channels; The intermediate vector is input into the Sigmoid activation function to obtain the channel attention weight map.

[0011] Secondly, this application provides a method for deploying an optimized model for fundus images, the method comprising: An initial fundus image optimization model trained in a server environment is obtained, wherein the initial fundus image optimization model is trained according to a fundus image optimization method provided in the first aspect as a framework; The initial fundus image optimization model is converted into an intermediate model in an intermediate representation format using a model conversion tool based on the Open Neural Network Exchange Standard. Configure the target inference engine in the system environment of the hardware platform where the fundus image optimization model is to be deployed; The intermediate model is loaded and parsed by the target inference engine to form an executable computation graph on the hardware platform, thereby completing the deployment of the fundus image optimization model.

[0012] Thirdly, this application provides a fundus imaging system, the system comprising: A fundus image capturing component is used to capture raw fundus images upon receiving a capturing command, and to send the captured raw fundus images to an image optimization component; An image optimization component, connected to the fundus image capturing component, is used to perform a fundus image optimization method as provided in the first aspect on the original fundus image to obtain an optimized target fundus image, and send the target fundus image to the image display component; An image display component is connected to both the image optimization component and the fundus image capturing component, and is used to display the original fundus image and / or the target fundus image.

[0013] Fourthly, this application provides a fundus imaging system, comprising: processor; Memory used to store the processor's executable instructions; The processor is configured to execute a fundus image optimization method as provided in the first aspect.

[0014] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: This application's embodiments acquire raw fundus images, perform frequency domain decomposition on the images, and optimize and reconstruct different components separately. This effectively enhances image details and suppresses noise, improving image contrast and clarity. Through layered upsampling combined with channel grouping and attention mechanisms for feature fusion, rich structural and textural details are accurately restored while increasing image resolution. Finally, by utilizing the synergistic enhancement of spatial and channel attention mechanisms, key diagnostically valuable regional features are adaptively highlighted, and background interference is further suppressed, ultimately generating a high-quality target fundus image with prominent details, appropriate contrast, and greater suitability for clinical observation and diagnosis. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic flowchart illustrating a fundus image optimization method provided in an embodiment of this application; Figure 2 This is an axonometric schematic diagram of the fundus imaging system provided in the embodiments of this application; Figure 3 This is a top cross-sectional view of the fundus imaging system provided in an embodiment of this application; Figure 4 A flowchart illustrating the model deployment method provided in this application embodiment; Figure 5 This is a schematic diagram of the architecture of the fundus imaging system provided in an embodiment of this application.

[0017] Figure label: 1-Optical lens, 2-Focus ring and aperture ring, 3-CMOS image sensor, 4-LED lamp bead, 5-Collimating lens, 6-Planar mirror, 7-20D lens, 8-Positioning light shield, 9-Touch screen, 10-Data processing module, 11-Housing. Detailed Implementation

[0018] This application provides a fundus image optimization method, a model deployment method, and a fundus imaging system, which solves the technical problem that fundus images acquired in the prior art have blurred details and insufficient contrast, resulting in low fundus image quality.

[0019] The technical solution of this application embodiment is to solve the above-mentioned technical problems, and the general idea is as follows: This application's embodiments acquire raw fundus images, perform frequency domain decomposition on the images, and optimize and reconstruct different components separately. This effectively enhances image details and suppresses noise, improving image contrast and clarity. Through layered upsampling combined with channel grouping and attention mechanisms for feature fusion, rich structural and textural details are accurately restored while increasing image resolution. Finally, by utilizing the synergistic enhancement of spatial and channel attention mechanisms, key diagnostically valuable regional features are adaptively highlighted, and background interference is further suppressed, ultimately generating a high-quality target fundus image with prominent details, appropriate contrast, and greater suitability for clinical observation and diagnosis.

[0020] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0021] First, it should be clarified that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0022] The eyes are an important window into the body's health, and the fundus is the only part of the body where blood vessels and nerves can be directly observed non-invasively. The clarity of fundus images directly affects the accuracy of early screening and diagnosis of various eye diseases (such as diabetic retinopathy, glaucoma, macular degeneration, etc.) and even systemic diseases. Therefore, obtaining high-quality fundus images is a crucial step in clinical ophthalmological examinations.

[0023] While traditional desktop fundus cameras offer high image quality, their bulky size, high cost, and complex operation limit their widespread application. To address this challenge, portable or handheld fundus imaging devices have emerged. However, existing portable devices still face key bottlenecks in their widespread adoption: limited by compact optical design, cost, and unstable shooting environments, these devices generally suffer from poor image quality, such as uneven brightness, low contrast, artifacts, and noise. This results in a large number of images being uninterpreted by doctors in clinical practice due to quality issues, severely impacting screening efficiency and diagnostic reliability.

[0024] To address these issues, related technologies offer two different approaches. One approach involves providing small, portable imaging devices with optimized hardware structures. These devices focus on optimizing optical structures and mechanical designs to achieve portability, but they do not integrate any advanced image processing or enhancement functions. Therefore, the output of these devices remains raw, inconsistent images, and the final imaging effect highly depends on the operator's skill and ideal shooting environment. In practical applications, the "usability" of the images remains low.

[0025] Another approach is to improve image quality through software algorithms, such as traditional image enhancement algorithms including histogram equalization, spatial filtering (IEEE TMI, 2018, 37, 2536), or frequency filtering (SignalProcessing, 2020, 170, 107445). However, these algorithms are usually based on fixed mathematical models, resulting in poor generalization ability, insufficient robustness and flexibility, and difficulty in handling the complex and varied low-quality scenes in fundus images, leading to limited enhancement effects. Another example is some deep learning enhancement algorithms; however, these algorithms fail to effectively recover multi-scale information from images. Furthermore, these deep learning enhancement algorithms typically contain millions or even hundreds of millions of parameters, requiring powerful computing capabilities, while the performance of embedded processors (MCU / CPU) in portable devices is limited, making it difficult to support real-time inference. These deep learning enhancement algorithms have high operating loads and high energy consumption, severely impacting device battery life and stability. Compressing and lightweighting these deep learning enhancement algorithms would reduce their accuracy, offering little improvement in fundus image quality.

[0026] To address the aforementioned problems, this application provides a method for optimizing fundus images, comprising steps S11-S14, as detailed below. Figure 1 As shown.

[0027] Step S11: Obtain the original fundus image to be optimized; Step S12: Optimize the multiple frequency domain feature components obtained after decomposing the original fundus image, and obtain a reconstructed feature image based on the optimized multiple frequency domain feature components; Step S13: Perform hierarchical upsampling on the reconstructed feature image, and obtain a high-resolution image through channel grouping, intra-group attention mechanism and cross-group feature fusion; Step S14: The high-resolution feature image is enhanced collaboratively based on spatial attention mechanism and channel attention mechanism to obtain an optimized target fundus image.

[0028] This application provides a method for optimizing fundus images, which can be executed by a fundus imaging system. The fundus imaging system may include a fundus image acquisition component, an image optimization component, and an image display component.

[0029] The fundus image capturing component is used to acquire raw fundus images upon receiving a capturing command, and to send the acquired raw fundus images to the image optimization component.

[0030] An image optimization component, connected to the fundus image capturing component, is used to execute a fundus image optimization method provided in this application embodiment on the original fundus image to obtain an optimized target fundus image, and then send the target fundus image to the image display component.

[0031] An image display component is connected to both the image optimization component and the fundus image capturing component, and is used to display the original fundus image and / or the target fundus image.

[0032] For example, the fundus imaging system provided in the embodiments of this application can refer to Figure 2 and Figure 3 As shown, the fundus image acquisition component includes an optical lens 1, a focus ring and an aperture ring 2, a CMOS image sensor 3, LED beads 4, a collimating lens 5, a plane mirror 6, a 20D lens 7, a positioning light shield 8, and a housing 11. The image optimization component includes a data processing module 10, and the image display component includes a touch screen 9.

[0033] A 20D lens 7 is used for front focusing, a plane mirror 6 is used to fold the optical path, the focal length of the optical lens 1 can be 16mm, and a CMOS image sensor 3 is used to convert optical signals into digital image signals. LED beads 4 provide the light source, and a collimating lens 5 converts the light emitted by the light source into uniform, shadowless light, providing illumination for fundus imaging. A touchscreen 9 is used for real-time viewing, parameter settings, and side-by-side comparison of fundus images before and after enhancement. The buttons displayed on the interface can be used to control shooting, function selection, and other operations. In addition, the fundus imaging system may also include a prompting unit and a rechargeable lithium battery. The prompting unit includes a built-in buzzer or voice module to provide operational feedback (such as prompts for successful focus or shooting completion), facilitating image capture by a single user. The built-in rechargeable lithium battery provides power to the entire system. The fundus imaging system is integrated into an ergonomically designed portable handheld housing 11, achieving system integration and portability.

[0034] Several LED beads 4 are arranged in a ring array to provide uniform illumination, and their brightness can be adjusted via system control. The light emitted by the LED beads 4 is collimated by a collimating lens 5 to form a parallel beam, then reflected by two plane mirrors 6, and finally illuminates the patient's fundus through a front-mounted 20D lens 7. The imaging light reflected from the fundus passes through the central area of ​​the 20D lens 7 and the plane mirrors 6, and enters the optical system inside the device. This light is then focused and imaged by the main lens group composed of optical lenses 1, and finally projected onto the photosensitive surface of the CMOS image sensor 3 to form a digital image. A focusing ring and an aperture ring 2 are provided at the front of the device, allowing the operator to manually perform precise focusing and depth-of-field control. At the same time, a positioning light shield 8 is located at the front of the device to assist the operator in determining the correct examination distance and effectively block external stray light from interfering with the imaging.

[0035] The data processing module 10 serves as the computing core of the system, with a high-performance embedded computing motherboard (e.g., Raspberry Pi 5) integrated inside the main body of the device. This module is responsible for receiving raw fundus images from the CMOS sensor 3 and performing optimization processing on the raw fundus images according to the fundus image optimization method provided in this embodiment. The touchscreen 9 provides an intuitive operating interface for the operator, integrating display and input functions. The operator can use the touchscreen 9 to perform real-time framing, parameter settings (such as LED brightness), trigger shooting, and browse and compare images before and after enhancement.

[0036] This application continues to incorporate the embodiments. Figure 1 A method for optimizing fundus images is described below.

[0037] Regarding step S11, obtain the original fundus image to be optimized.

[0038] Raw fundus images can be obtained by the fundus imaging component of a fundus imaging system. Alternatively, raw fundus images can be derived from existing medical image databases as input for offline processing.

[0039] The original fundus images may have problems such as insufficient resolution, low overall contrast, and blurred edges of key structures such as vascular networks and optic discs due to factors such as opacity of the refractive media of the patient's eye (such as cataracts), micro-tremors of the eyeball, inaccurate focusing of the equipment, or optical aberrations.

[0040] Regarding step S12, the multiple frequency domain feature components obtained after decomposing the original fundus image are optimized respectively, and a reconstructed feature image is obtained based on the optimized multiple frequency domain feature components.

[0041] Step S12 involves separating and processing the mixed features of different scales in the image. Specifically, frequency domain transformation (such as wavelet transform, contourlet transform, or non-subsampled contourlet transform NSCT) is first used to decompose the original fundus image into multiple frequency domain feature components carrying different information. These components typically include low-frequency components, mid-frequency components, and high-frequency components.

[0042] Low-frequency components mainly contain the overall appearance and background information of the image, but their contrast may be insufficient; mid-frequency components carry key structural information such as the main blood vessel walls and the outline of the optic disc; high-frequency components contain details such as the fine blood vessel ends, lesion edges and textures, but they are often mixed with a lot of noise.

[0043] Subsequently, differentiated optimizations were performed based on the characteristics of different components. Low-frequency components underwent contrast stretching to improve the overall visual effect; while mid- and high-frequency components employed enhancement algorithms based on nonlinear functions or adaptive gain strategies to sharpen structural edges, highlight texture details, and effectively suppress noise amplification. Finally, the optimized frequency domain feature components were reconstructed through corresponding inverse transforms, synthesizing a reconstructed feature image that retains all information from the original image while significantly enhancing contrast, sharpness, and detail, laying the foundation for subsequent super-resolution reconstruction.

[0044] Specifically, step S12 can be broken down into steps S1211-S1213.

[0045] Step S1211: Decompose the original fundus image to obtain multiple frequency domain feature components; Step S1212: Perform depthwise separable convolution processing on each frequency domain feature component to obtain each optimized frequency domain feature component. Step S1213: Reconstruct the optimized multiple frequency domain feature components to obtain the reconstructed feature image.

[0046] Step S1211 uses frequency domain transformation tools (such as wavelet transform) to separate the image into low-frequency, mid-frequency, and high-frequency components that carry different information. Step S1211 achieves accurate separation of the overall image contour, key structures, and subtle details / noise, creating conditions for subsequent targeted optimization.

[0047] Step S1212 employs depthwise separable convolution to process each component independently. Step S1212 achieves contrast enhancement for low-frequency components and detail sharpening and noise suppression for mid-to-high-frequency components with low computational cost, thereby efficiently completing the differentiated optimization of each component.

[0048] Step S1213 involves re-fusing the optimized frequency domain components through inverse transform. The reconstructed image, while preserving the original information, achieves a synergistic improvement in contrast, sharpness, and detail, forming a high-quality base image for subsequent steps.

[0049] In addition, step S12 can be further decomposed into the following steps S1221-S1223.

[0050] Step S1221: Decompose the original fundus image using wavelet transform to obtain multiple frequency domain feature components; the multiple frequency domain feature components include at least low-frequency approximation components, horizontal detail components, vertical detail components, and diagonal detail components. Step S1222: Perform depthwise separable convolution processing on each frequency domain feature component to obtain each optimized frequency domain feature component. Step S1223: The optimized multiple frequency domain feature components are reconstructed using inverse wavelet transform to obtain the reconstructed feature image.

[0051] Step S1221 aims to effectively separate global and local features in the image through multi-resolution analysis. Specifically, in step S1221, wavelet transform (such as Haar or Db4 wavelet) is used to decompose the original fundus image into four non-overlapping frequency domain sub-bands: low-frequency approximation component, horizontal detail component, vertical detail component, and diagonal detail component. The low-frequency approximation component carries the macroscopic structure and background illumination information of the image; the horizontal and vertical detail components respectively contain the longitudinal and transverse gradient information of structures such as blood vessel edges and optic disc contours; while the diagonal detail component reflects more high-frequency details such as texture and noise. Step S1221, through this lossless and orthogonal decomposition, provides a structurally clear and physically meaningful multi-scale feature basis for subsequent processing, which is a prerequisite for robust feature extraction.

[0052] Step S1222 involves customized feature refinement for different sub-bands. Step S1222 deploys depthwise separable convolutional modules in parallel for the four components. Each module first performs depthwise convolution to independently extract spatial features within each sub-band, capturing its unique directional patterns (such as the response of horizontal blood vessels in the horizontal detail component); subsequently, pointwise convolution is performed, using 1x1 convolutional kernels for cross-channel information integration and nonlinear transformation. Step S1222 enables the network to adaptively enhance discriminative features in each sub-band, for example, sharpening edges and suppressing irrelevant noise in the detail component, and optimizing contrast in the approximate component. Step S1222 achieves refined enhancement and denoising of multi-scale features with extremely low computational complexity, significantly improving feature discriminativity and laying the foundation for high-quality reconstruction.

[0053] Regarding step S1223, the four frequency domain feature components (i.e., the enhanced approximate, horizontal, vertical, and diagonal components) after depthwise separable convolution optimization are taken as input, and the inverse wavelet transform corresponding to step S1221 is performed. This transform, based on the wavelet basis function reconstruction algorithm, accurately fuses the four independent, enhanced sub-bands into a complete spatial image. Due to the orthogonality and perfect reconstruction characteristics of the wavelet transform, this process seamlessly integrates the optimization effects of all sub-bands, ensuring that macroscopic contrast, key structural contours, and microscopic texture details are synergistically enhanced in the output image. The final reconstructed feature image not only effectively preserves and highlights key structural details but also significantly improves overall visual quality, while ensuring the efficiency of the entire multi-scale feature extraction and optimization process.

[0054] Regarding step S13, the reconstructed feature image is upsampled hierarchically, and a high-resolution image is obtained through channel grouping, intra-group attention mechanism and cross-group feature fusion.

[0055] Step S13 involves converting the image optimized in step S12 from low resolution to high resolution. Specifically, the image size is first gradually increased through hierarchical upsampling. During this process, a channel grouping strategy is used to divide the feature map into different groups to focus on learning diverse feature representations. Subsequently, an attention mechanism is introduced within each group to autonomously strengthen the channel features that are crucial for image reconstruction. Finally, through cross-group feature fusion technology, the complementary features extracted from each group are effectively integrated, thereby accurately recovering rich details and structures while enlarging the image, ultimately generating a high-resolution image with clear details and accurate structure, laying the foundation for the final collaborative enhancement.

[0056] Specifically, step S13 may include steps S131-S135.

[0057] Step S131: The reconstructed feature image is subjected to hierarchical upsampling based on depth-separable convolution and pixel shuffling using multiple cascaded upsampling blocks to obtain an initial high-resolution image. Step S132: The initial high-resolution image is grouped into channels according to multiple preset channel feature characteristics to obtain feature sub-images corresponding to each channel feature characteristic. Step S133: Based on the response intensity of each channel in the current feature sub-map, determine the weight coefficient of each channel in each feature sub-map; Step S134: Based on the weight coefficient of each feature sub-map, perform feature recalibration on the corresponding feature sub-map to obtain the enhanced feature sub-map; Step S135: Perform cross-group feature fusion on each enhanced feature sub-image to obtain the high-resolution image.

[0058] Regarding step S131, multiple cascaded upsampling blocks refer to a group of modules connected in sequence specifically for enlarging the image size. Within each upsampling block, feature extraction and transformation are first performed using depthwise separable convolution, which effectively reduces computational complexity. Subsequently, a pixel shuffling operation is used to improve spatial resolution by reorganizing the channel information in the feature map, thereby achieving efficient scale enlargement. The initial high-resolution image is the output result obtained by progressively and layer-wise enlarging the reconstructed feature image through this series of upsampling blocks. It is much larger in scale than the input reconstructed feature image, but its structural and detail information needs further enhancement. Step S131 achieves efficient lightweight upsampling through depthwise separable convolution combined with channel rearrangement, gradually restoring low-resolution features to high-resolution space, thus expanding the image size while laying the foundation for subsequent fine processing.

[0059] Regarding step S132, in image processing, a channel typically refers to the feature representation of a feature map in different dimensions. Each channel can be considered as extracting a specific image characteristic (such as edge, texture, or color). Multiple predefined channel feature characteristics refer to predefined grouping criteria used to distinguish different semantic information categories. For example, they can be categorized based on the type of information they carry, such as structural features, texture features, and global context features. Channel grouping involves dividing all channels of the initial high-resolution image into different subsets based on these predefined characteristics. Each feature sub-map is a set of channels obtained after grouping, focusing on a specific type of feature characteristic. These represent the specific representation of the initial high-resolution image from different feature perspectives. Step S132, by grouping the initial high-resolution image by channel, creates conditions for the subsequent intra-group attention mechanism, enabling the model to group and model different types of features, thereby processing multi-scale feature embeddings more precisely.

[0060] Regarding step S133, the response intensity of each channel in the feature sub-map refers to the overall activity or importance level reflected by the activation values ​​of that channel at all spatial locations. A high response intensity usually means that the channel carries more critical feature information. The weight coefficient of each channel is a scalar value between 0 and 1, and its magnitude is positively correlated with the response intensity of the channel. Its function is to quantify the relative importance of the channel in the current feature sub-map, so as to enhance important features and suppress secondary features in subsequent steps. Step S133 determines its weight coefficient by analyzing the response intensity of each channel in each feature sub-map, realizing dynamic weight allocation, providing data support for subsequent accurate enhancement of key features, and is a key step in strengthening the expression of key structural and texture information.

[0061] Regarding step S134, feature recalibration of the feature sub-map based on weight coefficients works by amplifying the contribution of important feature channels and weakening the influence of secondary or noisy channels through channel-level multiplication operations, based on the importance of each channel (i.e., weight coefficient). Specifically, each channel of the feature sub-map is multiplied by its corresponding weight coefficient to obtain a recalibrated and enhanced feature sub-map. Step S134, by recalibrating the corresponding feature sub-map using weight coefficients, significantly strengthens the expression of key structural and texture information crucial for image reconstruction, effectively improving feature quality and enhancing the model's ability to capture and retain important information.

[0062] Regarding step S135, the principle of cross-group feature fusion refers to the reintegration of complementary feature information (such as structure, texture, and context) carried by each feature sub-image after grouping and independent enhancement. This process is usually achieved through methods such as channel concatenation followed by convolutional transformation, aiming to generate a complete and unified enhanced feature representation. Step S135 performs cross-group feature fusion on each enhanced feature sub-image, effectively integrating multi-scale feature embeddings from different groups, resulting in a high-resolution image with both clear structure and rich detail. Compared to the input reconstructed feature image, the final high-resolution image is not only larger in spatial size, but more importantly, it significantly improves structural integrity and detail richness, providing higher-quality and more easily interpreted image evidence for final clinical diagnosis.

[0063] Regarding step S14, the high-resolution feature image is enhanced collaboratively based on spatial attention mechanism and channel attention mechanism to obtain an optimized target fundus image.

[0064] Step S14 aims to further improve the visual quality and diagnostic value of high-resolution images. Specifically, firstly, a channel attention mechanism adaptively calibrates the weights of each feature channel, emphasizing information-rich feature channels and suppressing redundant channels, thereby enhancing the model's ability to discriminate key pathological features. Simultaneously, a spatial attention mechanism focuses on spatial relationships within the image, dynamically increasing the pixel weights of key structures such as vessel edges and lesion areas to sharpen details and improve spatial contrast. Ultimately, through the synergistic effect of these two attention mechanisms, the model can jointly optimize the image from both "feature channels" and "spatial location" dimensions, generating a target fundus image with prominent details, appropriate contrast, and greater suitability for clinical observation and diagnosis.

[0065] Specifically, step S14 may include steps S141-S144.

[0066] Step S141: Based on the spatial attention mechanism, determine the spatial attention weight map corresponding to the high-resolution feature image; Step S142: Based on the channel attention mechanism, determine the channel attention weight map corresponding to the high-resolution feature image; Step S143: The spatial attention weight map and the channel attention weight map are fused to obtain the fused attention weight; Step S144: The high-resolution feature image is calibrated based on the fused attention weights to obtain the target fundus image.

[0067] Regarding step S141, firstly, max pooling is performed on the high-resolution feature image along the channel dimension to obtain a salient feature map; then, average pooling is performed on the high-resolution feature image along the channel dimension to obtain an average feature map; subsequently, the salient feature map and the average feature map are concatenated along the channel dimension to obtain a fused feature map; finally, the fused feature map is sequentially input into a standard convolutional layer and a sigmoid activation function to obtain a spatial attention weight map.

[0068] Max pooling extracts the most salient response at each location across all channels in the spatial dimension, resulting in a salient feature map that highlights the most striking local features in the image. Average pooling, on the other hand, calculates the average response at each location across all channels, resulting in an average feature map that reflects the overall background and contextual information of the image. The salient feature map focuses more on highlighting local extrema (such as possible lesion edges), while the average feature map preserves a smoother overall structure; the two are complementary in their information emphasis. Subsequently, these two feature maps are concatenated along the channel dimension, fusing complementary spatial information from "salient features" and "average background" to form a more informative fused feature map. This fused feature map is then input into a standard convolutional layer for feature compression and integration, and the output value is normalized to between 0 and 1 using a sigmoid activation function, ultimately generating a spatial attention weight map. The principle behind this weight map is to assign a weight value to each spatial location in the high-resolution feature image, thereby indicating the importance of that location. Step S141 performs max pooling and average pooling along the channel dimension on the high-resolution feature image, and generates a spatial attention weight map through convolutional layers and the Sigmoid function, so that the model can focus on the spatial saliency of key pathological areas such as the optic disc and microaneurysms, effectively improving the visual recognition of vascular contours and small lesions.

[0069] Regarding step S142, firstly, global average pooling is performed on the high-resolution feature image to obtain channel statistical vectors; then, the channel statistical vectors are input into a bottleneck structure neural network to obtain intermediate vectors representing the dependencies between channels; finally, the intermediate vectors are input into a sigmoid activation function to obtain a channel attention weight map.

[0070] Global average pooling averages the values ​​at all spatial locations of each feature channel to obtain a one-dimensional channel statistical vector. Each element of this vector summarizes the overall activation intensity of a feature channel. This channel statistical vector is then input into a bottleneck-structured neural network (typically composed of two fully connected layers with a lower-dimensional intermediate layer to form a bottleneck). This network effectively models and captures the complex nonlinear dependencies between channels, and its output intermediate vector encodes these relationships. This intermediate vector is then processed by the Sigmoid activation function to normalize the weights of each channel to between 0 and 1, ultimately yielding a channel attention weight map. The working principle of this weight map is to adaptively assign a weight coefficient to each feature channel based on its relevance to the diagnostic task. Step S142, by performing global average pooling on the high-resolution feature image and generating a channel attention weight map via a bottleneck-structured neural network and the Sigmoid function, can adaptively strengthen feature channels highly correlated with diagnostic information such as texture and edges, while suppressing redundant or noise-dominated channel responses.

[0071] Regarding step S143, the spatial attention weight map and the channel attention weight map are fused. The principle is to combine the importance weights calculated separately from the "spatial dimension" and the "channel dimension," commonly using element-wise addition or multiplication to obtain a fused attention weight that simultaneously reflects the importance of spatial location and channel features. The core function of this fused weight is to construct a unified view, collaboratively guiding the model to prioritize enhancements "where" (space) and "what features" (channels). Step S143, by fusing the spatial and channel attention weight maps, achieves the synergistic effect of spatial and channel attention mechanisms, laying the foundation for ultimately generating images that simultaneously enhance diagnostically valuable key regions and suppress interference.

[0072] Regarding step S144, the process of calibrating the high-resolution feature image based on fused attention weights essentially involves multiplying the fused weights element-wise with the original high-resolution feature map. This operation is equivalent to a fine filter, simultaneously amplifying or suppressing the feature map in both spatial and channel dimensions based on the importance determined by the fused weights. The resulting target fundus image, compared to the input high-resolution feature image, demonstrates significant improvements and advantages: not only is it higher resolution, but more importantly, through collaborative calibration using the attention mechanism, the feature representation of key diagnostic regions (such as fine blood vessels and lesions) is significantly enhanced, while background noise and artifact interference are effectively suppressed, resulting in a qualitative leap in visual recognition and diagnostic value. Step S144 calibrates the high-resolution feature image based on fused attention weights, ultimately generating an optimized target fundus image that effectively enhances the feature representation of key diagnostic regions while suppressing noise and artifact interference, greatly improving the image's usability and reliability for clinical diagnosis.

[0073] In summary, this embodiment of the application, by acquiring the original fundus image, performing frequency domain decomposition on the image, and optimizing and reconstructing different components, effectively enhances image details and suppresses noise, thereby improving image contrast and clarity. Through hierarchical upsampling combined with channel grouping and attention mechanisms for feature fusion, rich structural and textural details are accurately restored while amplifying image resolution. Finally, by utilizing the synergistic enhancement of spatial and channel attention mechanisms, key features with diagnostic value are adaptively highlighted, and background interference is further suppressed, ultimately generating a high-quality target fundus image with prominent details, appropriate contrast, and greater suitability for clinical observation and diagnosis.

[0074] First, addressing the aforementioned issues of poor generalization ability and difficulty in handling complex, low-quality scenarios in software algorithms, step S12 of this application embodiment decomposes the original image in the frequency domain and optimizes different components separately. This achieves differentiated and adaptive processing of image details, structure, and noise, significantly improving the algorithm's robustness and enhancement effect against variable degradation factors. Second, addressing the aforementioned bottlenecks of large parameter count, high computational load, and difficulty in deploying deep learning algorithms on portable devices, the hierarchical upsampling, channel grouping, and intra-group attention mechanisms used in step S13 of this application embodiment, along with the depthwise separable convolution used in step S12, together constitute a highly efficient and lightweight network structure. This significantly reduces model complexity and computational energy consumption while ensuring performance, making it more suitable for running on resource-constrained embedded processors. Furthermore, addressing the aforementioned issue that related deep learning algorithms failed to effectively recover multi-scale information in images, the frequency domain decomposition and optimization in step S12 and the hierarchical upsampling and cross-group feature fusion in step S13 of this application embodiment collaboratively achieve effective extraction and reconstruction of multi-scale features from low-frequency contours to high-frequency details, ensuring that the final image possesses rich structural information and detailed levels. Finally, through the synergistic enhancement of the spatial and channel attention mechanisms in step S14 of this application embodiment, key diagnostic regions are further highlighted and interference is suppressed, thereby significantly improving the "availability" of the output image at the software level and reducing excessive reliance on hardware performance and operator skills.

[0075] Regarding image quality, the embodiments of this application effectively improve the problems of low clarity, poor contrast, and insufficient signal-to-noise ratio commonly found in the original images through frequency domain decomposition and targeted optimization in step S12. The hierarchical upsampling and feature fusion mechanism in step S13 further restores the multi-scale details and structural information of the image. Finally, through collaborative attention enhancement in step S14, key pathological features are adaptively sharpened and background interference is suppressed. This enhances the originally inconsistent quality of the original images into high-quality images with prominent details and distinct features, greatly improving their clinical usability and diagnostic reliability.

[0076] Secondly, regarding algorithm efficiency, the depthwise separable convolution (step S1212), channel grouping, and efficient attention mechanism (steps S13 and S14) employed in this embodiment constitute a lightweight network architecture with low computational load. This design significantly reduces the number of model parameters and computational complexity, enabling it to run efficiently on resource-constrained embedded processors (MCU / CPU) with the help of optimization techniques such as model quantization. This achieves high-quality real-time image enhancement on portable devices, breaking the balance between performance and power consumption.

[0077] Third, at the system integration and application level, this application embodiment deeply integrates the aforementioned efficient software algorithm (steps S11-S14) with portable hardware, providing users with a simple, "instant" integrated solution. This highly integrated design not only significantly reduces the system's reliance on operator expertise and usage costs, but also powerfully promotes the widespread adoption of high-quality fundus examination technology in primary healthcare and large-scale screening scenarios, improving the accessibility and efficiency of early diagnosis of ophthalmic diseases.

[0078] Based on the same inventive concept, the embodiments of this application provide, as follows: Figure 4 The method shown is a fundus image optimization model deployment method, which includes steps S41-S44.

[0079] Step S41: Obtain the initial fundus image optimization model trained in the server environment. The initial fundus image optimization model is trained according to the fundus image optimization method provided above. Step S42: The initial fundus image optimization model is converted into an intermediate model in an intermediate representation format using a model conversion tool based on the Open Neural Network Exchange Standard. Step S43: Configure the target inference engine in the system environment of the hardware platform where the fundus image optimization model is to be deployed; Step S44: Load and parse the intermediate model through the target inference engine to form an executable computation graph on the hardware platform, and complete the deployment of the fundus image optimization model.

[0080] Step S41 is crucial in obtaining the initial fundus image optimization model trained in a high-performance server environment. This model, based on the "fundus image optimization method" described in the claims, is trained extensively using a large-scale fundus image dataset and possesses powerful image enhancement capabilities. This step ensures that the model deployed to the terminal device has proven high-precision performance, laying a solid foundation for maintaining excellent enhancement effects in resource-constrained environments. Step S41 separates the complex model training process from the lightweight deployment application, guaranteeing the core competitiveness of the final deployed model.

[0081] Regarding step S42, a model conversion tool based on the Open Neural Network Exchange (ONNX) standard is used to convert the initial model from a specific training framework (such as PyTorch) into a hardware-platform-independent intermediate representation format. This conversion process unifies the model description and effectively decouples the model training framework from the terminal deployment environment. Step S42 solves the compatibility issues between different frameworks and hardware platforms, providing unprecedented convenience and standardized support for the subsequent flexible and efficient deployment of models on diverse embedded hardware.

[0082] Regarding step S43, environmental preparation is performed for the efficient operation of the model on the terminal hardware. Step S43 pre-configures a target inference engine (e.g., TensorFlow Lite or ONNX Runtime) specifically optimized for the target hardware platform (e.g., the embedded processor of a portable fundus camera) within the system environment of that hardware platform. This inference engine encapsulates low-level computation libraries and acceleration instructions (e.g., NEON SIMD) customized for a specific chip architecture (e.g., an ARM CPU). Through hardware-software co-optimization, step S43 maximizes the limited computational potential of portable devices, providing a crucial software runtime environment for subsequent high-speed, low-power inference of the model.

[0083] Regarding step S44, the target inference engine deployed on the hardware platform loads and parses the intermediate model generated in step S42, transforming it into a computational graph that the hardware platform can directly understand and execute. This process realizes the mapping of the deep learning model from a file to hardware-executable instructions. Step S44 ultimately successfully forms an efficient retinal image optimization instance on a resource-constrained portable device, enabling the high-precision algorithm trained on the server to seamlessly complete real-time inference on the terminal, truly achieving the ultimate goal of "instantly obtaining" high-quality retinal images.

[0084] Based on the same inventive concept, embodiments of this application provide a fundus imaging system, the system comprising: A fundus image capturing component is used to capture raw fundus images upon receiving a capturing command, and to send the captured raw fundus images to an image optimization component; An image optimization component, connected to the fundus image capturing component, is used to perform a fundus image optimization method as described above on the original fundus image to obtain an optimized target fundus image, and send the target fundus image to the image display component; An image display component is connected to both the image optimization component and the fundus image capturing component, and is used to display the original fundus image and / or the target fundus image.

[0085] Based on the same inventive concept, the embodiments of this application provide, as follows:Figure 5 The fundus imaging system shown includes: Processor 51; Memory 52 for storing executable instructions of the processor 51; The processor 51 is configured to execute a fundus image optimization method as described above.

[0086] The present application provides a comprehensive example to illustrate the process of optimizing fundus images according to the embodiments of the present application.

[0087] The operator turns on the power to the fundus imaging system. After the fundus imaging system is started, the touchscreen 9 enters live view mode, and the operator can aim the device at the patient's eye and focus based on the live image displayed on the screen.

[0088] Once the image is focused, the operator presses the shutter button on the touchscreen 9. The image acquisition module then captures one or more frames of raw fundus images and temporarily stores them in memory.

[0089] After acquisition, the fundus imaging system automatically sends the raw images to the internal data processing module 10 without user intervention. The fundus image enhancement unit within the module executes one of the aforementioned fundus image optimization methods locally, generating a corresponding enhanced image within seconds. This process is entirely offline and does not rely on any external network connection.

[0090] After the enhancement process is complete, the touchscreen 9's interface will automatically switch, displaying the original image captured and the high-quality enhanced image obtained after processing in split-screen or side-by-side. This instant comparison greatly helps the operator judge the effect of image quality improvement and assists in preliminary lesion observation.

[0091] The operator can select to save the image via touchscreen 9. The fundus imaging system will store both the original and enhanced images on a local memory card. Subsequently, the images can be exported to devices such as computers via Wi-Fi for the creation of medical records or further analysis.

[0092] Since the electronic device described in this embodiment is an electronic device used to implement the information processing method in the embodiments of this application, those skilled in the art can understand the specific implementation methods and various variations of the electronic device in this embodiment based on the information processing method described in the embodiments of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. Any electronic device used by those skilled in the art to implement the information processing method in the embodiments of this application falls within the scope of protection of this application.

[0093] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0094] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0095] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0097] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0098] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for optimizing fundus images, characterized in that, The method includes: Obtain the original fundus image to be optimized; The multiple frequency domain feature components obtained after decomposing the original fundus image are optimized respectively, and a reconstructed feature image is obtained based on the optimized multiple frequency domain feature components. The reconstructed feature image is upsampled hierarchically, and a high-resolution image is obtained through channel grouping, intra-group attention mechanism and cross-group feature fusion; The high-resolution feature image is enhanced collaboratively using spatial attention and channel attention mechanisms to obtain an optimized target fundus image.

2. The fundus image optimization method as described in claim 1, characterized in that, The multiple frequency domain feature components obtained after decomposing the original fundus image are optimized respectively, and a reconstructed feature image is obtained based on the optimized multiple frequency domain feature components, including: The original fundus image is decomposed to obtain multiple frequency domain feature components; Each frequency domain feature component is subjected to depthwise separable convolution to obtain the optimized frequency domain feature components. The optimized multiple frequency domain feature components are reconstructed to obtain the reconstructed feature image.

3. The fundus image optimization method as described in claim 1, characterized in that, The multiple frequency domain feature components obtained after decomposing the original fundus image are optimized respectively, and a reconstructed feature image is obtained based on the optimized multiple frequency domain feature components, including: The original fundus image is decomposed using wavelet transform to obtain multiple frequency domain feature components; the multiple frequency domain feature components include at least low-frequency approximation components, horizontal detail components, vertical detail components, and diagonal detail components; Each frequency domain feature component is subjected to depthwise separable convolution to obtain the optimized frequency domain feature components. The reconstructed feature image is obtained by reconstructing multiple frequency domain feature components using inverse wavelet transform.

4. The fundus image optimization method as described in claim 1, characterized in that, The reconstructed feature image is subjected to hierarchical upsampling, and a high-resolution image is obtained through channel grouping, intra-group attention mechanism, and cross-group feature fusion, including: Multiple cascaded upsampling blocks are used to perform hierarchical upsampling on the reconstructed feature image based on depth-separable convolution and pixel shuffling to obtain an initial high-resolution image; The initial high-resolution image is grouped into channels according to multiple preset channel features to obtain feature sub-images corresponding to each channel feature. Based on the response intensity of each channel in the current feature map, determine the weight coefficient of each channel in each feature map; Based on the weight coefficients of each feature sub-map, the corresponding feature sub-map is recalibrated to obtain the enhanced feature sub-map; Cross-group feature fusion is performed on each enhanced feature sub-image to obtain the high-resolution image.

5. The fundus image optimization method as described in claim 1, characterized in that, The high-resolution feature image is synergistically enhanced based on spatial attention and channel attention mechanisms to obtain an optimized target fundus image, including: Based on the spatial attention mechanism, the spatial attention weight map corresponding to the high-resolution feature image is determined; Based on the channel attention mechanism, the channel attention weight map corresponding to the high-resolution feature image is determined; The spatial attention weight map and the channel attention weight map are fused to obtain the fused attention weight; The high-resolution feature image is calibrated based on the fused attention weights to obtain the target fundus image.

6. The fundus image optimization method as described in claim 5, characterized in that, Based on the spatial attention mechanism, the spatial attention weight map corresponding to the high-resolution feature image is determined, including: Max pooling along the channel dimension is performed on the high-resolution feature image to obtain a salient feature map; The high-resolution feature image is subjected to average pooling along the channel dimension to obtain an average feature map; The salient feature map and the average feature map are concatenated along the channel dimension to obtain a fused feature map; The fused feature map is sequentially input into a standard convolutional layer and a Sigmoid activation function to obtain a spatial attention weight map.

7. The fundus image optimization method as described in claim 5, characterized in that, Based on the channel attention mechanism, the channel attention weight map corresponding to the high-resolution feature image is determined, including: Global average pooling is performed on the high-resolution feature image to obtain the channel statistics vector; The channel statistical vectors are input into the bottleneck structure neural network to obtain the intermediate vectors representing the dependencies between channels; The intermediate vector is input into the Sigmoid activation function to obtain the channel attention weight map.

8. A method for deploying an optimized fundus image model, characterized in that, The method includes: Obtain an initial fundus image optimization model trained in a server environment, wherein the initial fundus image optimization model is trained according to a fundus image optimization method as described in any one of claims 1-7; The initial fundus image optimization model is converted into an intermediate model in an intermediate representation format using a model conversion tool based on the Open Neural Network Exchange Standard. Configure the target inference engine in the system environment of the hardware platform where the fundus image optimization model is to be deployed; The intermediate model is loaded and parsed by the target inference engine to form an executable computation graph on the hardware platform, thereby completing the deployment of the fundus image optimization model.

9. A fundus imaging system, characterized in that, The system includes: A fundus image capturing component is used to capture raw fundus images upon receiving a capturing command, and to send the captured raw fundus images to an image optimization component; An image optimization component, connected to the fundus image capturing component, is used to perform a fundus image optimization method as described in any one of claims 1-7 on the original fundus image to obtain an optimized target fundus image, and send the target fundus image to the image display component; An image display component is connected to both the image optimization component and the fundus image capturing component, and is used to display the original fundus image and / or the target fundus image.

10. A fundus imaging system, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute a fundus image optimization method as described in any one of claims 1 to 7.

Images