Portable keratitis real-time screening device based on lightweight edge computing

Abstract

This invention provides a portable real-time keratitis screening device based on lightweight edge computing, belonging to the fields of medical auxiliary diagnostic equipment and computer vision technology. The device includes an optical acquisition unit, a control unit, and an interaction unit. The interaction unit receives image acquisition commands. The optical acquisition unit includes a macro camera module, which, based on the image acquisition commands, calls the macro camera module to acquire eye images of the target object. The control unit includes an edge development board, on which a lightweight image recognition algorithm is deployed on a neural processing unit hardware accelerator. The control unit performs real-time keratitis screening on the eye images based on the lightweight image recognition algorithm to obtain the keratitis screening results for the target object. This invention significantly improves inference efficiency, achieving millisecond-level keratitis screening, blocking intercepted or tampered paths, thereby mitigating the risk of data leakage.

Description

Technical Field

[0001] This invention relates to the fields of medical auxiliary diagnostic equipment and computer vision technology, and in particular to a portable real-time keratitis screening device based on lightweight edge computing. Background Technology

[0002] Keratitis is one of the leading causes of vision impairment and blindness worldwide, posing a serious threat to human visual health. Undiagnosed keratitis can easily lead to corneal ulcers and scarring, resulting in irreversible vision damage or even blindness.

[0003] Existing intelligent diagnostic solutions for keratitis typically process the entire slit-lamp image directly or operate on cropped local areas of the cornea or conjunctiva. While these methods achieve intelligent diagnosis of keratitis, the mainstream technical approach in current ophthalmic AI-assisted diagnostic systems heavily relies on a centralized cloud computing architecture.

[0004] While centralized architectures offer computational advantages in processing massive amounts of medical image data and running complex, demanding models, and have provided early technical validation for AI-based keratitis diagnosis, their practical application in clinical screening suffers from drawbacks such as high latency, vulnerability to privacy breaches, inability to function offline, and high deployment costs. These shortcomings have become significant technical barriers hindering their wider clinical application and promotion. Therefore, developing a novel architecture capable of breaking free from cloud constraints and achieving high-precision real-time inference on resource-constrained edge devices has become an inevitable trend in industry evolution. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a portable real-time keratitis screening device based on lightweight edge computing.

[0006] In a first aspect, the present invention provides a portable real-time keratitis screening device based on lightweight edge computing, comprising: an optical acquisition unit, a control unit, and an interaction unit; The interactive unit is used to receive image acquisition commands; The optical acquisition unit includes a macro camera module, and the optical acquisition unit is used to call the macro camera module to acquire an image of the target object's eye based on the image acquisition command; The control unit includes an edge development board, on which a lightweight image recognition algorithm is deployed on the neural processing unit hardware accelerator. The control unit is used to perform real-time keratitis screening processing on the eye image based on the lightweight image recognition algorithm to obtain the keratitis screening result of the target object.

[0007] In some embodiments, the optical acquisition unit further includes a time-of-flight sensor module; The optical acquisition unit is also used to acquire the flight time of the light pulse between the corneal surface of the target object and the macro camera module through the time-of-flight sensor module; The control unit is also used to calculate the distance between the corneal surface and the macro camera module based on the flight time, and generate a capture command when the distance is within the macro imaging range; The optical acquisition unit is also used to respond to the capture command by calling the macro camera module to automatically focus and acquire the eye image.

[0008] In some of these embodiments, it also includes: Battery management module; The battery management module includes a battery pack, a charging module for energy management of the battery pack, and a step-down converter module. The battery management module is used to convert the voltage output by the battery pack into a stable DC voltage through the step-down converter module, and to power the optical acquisition unit, the control unit and the interaction unit through the DC voltage.

[0009] In some of these embodiments, the buck converter module includes a synchronous buck converter, a transient voltage suppressor, a multi-stage decoupling filter capacitor, an energy storage inductor, and an output filter capacitor bank. The transient voltage suppressor and the multi-stage decoupling filter capacitor are respectively connected to the input terminal of the synchronous buck converter; The energy storage inductor is connected to the output terminal of the synchronous buck converter; The current corresponding to the voltage output by the battery pack flows into the synchronous buck converter for buck conversion through the transient voltage suppressor and the multi-stage decoupling filter capacitor to obtain the DC voltage. When the internal high-side power transistor of the synchronous buck converter is turned on, the current corresponding to the DC voltage flows into the optical acquisition unit, the control unit, and the interaction unit through the energy storage inductor.

[0010] In some embodiments, the buck converter module further includes a pull-up resistor and an indicator circuit; The pull-up resistor is connected to the output terminal of the synchronous buck converter and is used to output the status information of the synchronous buck converter; The indicator circuit is used to display the status information of the synchronous buck converter through a two-stage LED.

[0011] In some embodiments, the lightweight image recognition algorithm is a lightweight version of the MobileNetV4 algorithm; The lightweight MobileNetV4 algorithm is the RKNN format MobileNetV4 algorithm; The RKNN format MobileNetV4 algorithm is obtained by deeply optimizing the ONNX format MobileNetV4 algorithm using the RKNN-Toolkit toolchain; The ONNX format MobileNetV4 algorithm is obtained by converting the PTH format MobileNetV4 algorithm trained with PyTorch.

[0012] In some of these embodiments, the control unit is specifically used for: The eye image is cropped and standardized to obtain a corneal image; The corneal image is input into the lightweight image recognition algorithm, and the features of the corneal image are extracted by a general inverted residual block. The feature extraction results are analyzed to obtain the keratitis screening results, which are classification probability matrices for three categories: keratitis, normal, and other corneal abnormalities.

[0013] In some of these embodiments, the interaction unit includes a touch display screen; The touch display screen is used to generate the image acquisition command in response to the user's touch operation; Or, the touch display screen is used to display the eye image and / or the keratitis screening result.

[0014] In some of these embodiments, the control unit is specifically used for: Multiplying the flight time, speed of light, and preset values ​​yields the distance between the corneal surface and the macro camera module.

[0015] In some embodiments, the macro camera module includes a macro camera and a supplementary lighting device disposed around the macro camera; The supplemental lighting device is used to adaptively adjust the brightness according to the intensity of ambient light; The control unit is connected to the interaction unit via a mobile industrial processor interface.

[0016] Secondly, the present invention also provides a portable real-time keratitis screening method based on lightweight edge computing, applied to a portable real-time keratitis screening device based on lightweight edge computing. The portable real-time keratitis screening device based on lightweight edge computing includes an optical acquisition unit, a control unit, and an interaction unit. The optical acquisition unit includes a macro camera module, and the control unit includes an edge development board. A lightweight image recognition algorithm is deployed on the neural processing unit hardware accelerator of the edge development board. The method includes: The interactive unit receives image acquisition commands. The optical acquisition unit, based on the image acquisition command, calls the macro camera module to acquire an image of the target object's eye. The control unit performs real-time keratitis screening on the eye image based on the lightweight image recognition algorithm to obtain the keratitis screening result of the target object.

[0017] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to implement the portable real-time keratitis screening method based on lightweight edge computing described in the second aspect above.

[0018] In a fourth aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the portable real-time keratitis screening method based on lightweight edge computing described in the second aspect above.

[0019] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the portable real-time keratitis screening method based on lightweight edge computing described in the second aspect above.

[0020] Compared with existing technologies, the portable real-time keratitis screening device based on lightweight edge computing provided by this invention significantly improves inference efficiency by deploying a lightweight image recognition algorithm on the neural processing unit hardware accelerator of the edge development board. By realizing a closed loop from framing and preprocessing to inference result output locally on the device, the response gap caused by network transmission is completely eliminated, achieving millisecond-level "what you see is what you get" real-time framing assistance, which greatly improves the efficiency of doctors in capturing subtle lesions. Through the lightweight image recognition algorithm, high-performance image feature expression and classification are achieved with a very small number of parameters, so that complex medical image recognition tasks do not need to stream private data over the Internet, but complete the entire cycle of calculation within the terminal device's memory and neural processing unit. This data processing mechanism that does not leave the device physically blocks the path of interception or tampering, solving the risk of leakage.

[0021] Details of one or more embodiments of the present invention are set forth in the following drawings and description to make other features, objects and advantages of the invention more readily apparent. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of the invention and constitute a part of this invention, illustrate the invention. Those skilled in the art will recognize that other drawings can be derived from these drawings without any inventive effort. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is one of the structural schematic diagrams of the portable real-time keratitis screening device based on lightweight edge computing provided by the present invention.

[0023] Figure 2 This is the second schematic diagram of the structure of the portable real-time keratitis screening device based on lightweight edge computing provided by the present invention.

[0024] Figure 3 This is a physical schematic diagram of the time-of-flight sensor module provided by the present invention.

[0025] Figure 4 This is a circuit diagram of the buck converter module provided by the present invention.

[0026] Figure 5 This is a schematic diagram of the process for deploying the MobileNetV4 algorithm provided in an embodiment of the present invention.

[0027] Figure 6 This is a schematic diagram of the interactive interface provided by the present invention.

[0028] Figure 7This is a flowchart illustrating the portable real-time keratitis screening method based on lightweight edge computing provided by the present invention.

[0029] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0030] To more clearly understand the objectives, technical solutions, and advantages of this invention, the technical solutions of this invention will be clearly and completely described and explained below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0031] Unless otherwise defined, the technical or scientific terms used in this invention shall have the general meaning understood by one of ordinary skill in the art to which this invention pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this invention do not indicate quantitative limitation and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this invention are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” and “coupled” used in this invention are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “A plurality” used in this invention refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. The terms "first," "second," "third," etc., used in this invention are merely for distinguishing similar objects and do not represent a specific ordering of the objects.

[0032] First, a brief description of the relevant content involved in this invention will be given.

[0033] Currently, slit-lamp examination is the "gold standard" for diagnosing corneal lesions, allowing specialists to assess the nature and extent of the lesions by observing slit-lamp images. However, traditional slit-lamp equipment is bulky and expensive, primarily deployed in large general hospitals or specialized ophthalmology hospitals, and the diagnostic process heavily relies on the clinical experience of ophthalmologists. For many grassroots medical institutions or remote areas, the lack of expensive, sophisticated equipment and experienced doctors often prevents patients from receiving timely early screening, leading to delays in diagnosis and treatment. Therefore, developing an efficient, accurate, and lightweight method for corneal inflammation screening is a pressing issue that needs to be addressed.

[0034] With the rapid development of deep learning algorithms and ophthalmic image data, the application of AI-based slit-lamp and ocular surface image processing in intelligent diagnosis of eye diseases is gradually becoming a reality. Deep learning models can automatically extract features from a large number of ophthalmic images, achieving rapid and accurate intelligent diagnosis of eye diseases. In recent years, many researchers have proposed various keratitis diagnosis methods based on deep learning and slit-lamp images. For example, in 2021, researchers proposed an automatic keratitis screening algorithm based on a Convolutional Neural Network (CNN) algorithm and verified its effectiveness on original slit-lamp images from different hospitals. Based on this, in 2023, researchers proposed an automatic keratitis screening algorithm based on deep learning target localization. The target localization algorithm is used to automatically crop the region of interest (ROI) of the cornea and conjunctiva from the slit-lamp image, filter out noise and irrelevant features around the cornea and conjunctiva, and input this ROI into a cost-sensitive attention CNN to classify normal cornea, keratitis, and other corneal abnormalities, improving the accuracy and robustness of intelligent keratitis diagnosis. In 2024, other researchers proposed an automatic diagnostic algorithm based on hierarchical optimization technology. By using a hierarchical loss function to simulate the diagnostic process of ophthalmologists, the algorithm was able to classify normal cornea, bacterial keratitis, fungal keratitis, viral keratitis, and amoebic keratitis, further promoting the application of artificial intelligence algorithms in the clinical diagnosis of keratitis.

[0035] The aforementioned solutions primarily utilize digital slit lamp equipment to acquire images of the patient's eyes. The high-resolution raw images are then uploaded to a high-performance cloud server cluster via a wireless network. Subsequently, a deep convolutional neural network visual model deployed in the cloud is invoked to perform feature extraction and lesion classification. Finally, the diagnostic results are sent to the client.

[0036] For example, current solutions for intelligent diagnosis of keratitis based on slit-lamp images include: Option 1: Intelligent Diagnosis of Keratitis Based on Whole-Image Slit-Lamp Images: This method directly processes raw or pre-processed whole-image slit-lamp images, using machine learning algorithms or deep learning to extract high-level features from the slit-lamp images and classify keratitis. For keratitis screening, the algorithm automatically learns from the whole slit-lamp image and extracts key structural features, such as corneal transparency, morphology, and color changes. These features are then input into a classifier to distinguish between normal cornea, keratitis, and other corneal abnormalities. Furthermore, deep learning based on slit-lamp images can automatically diagnose early, middle, and late stages of keratitis by recognizing changes in lesion morphology and color.

[0037] Option 2: Automatic Localization and Intelligent Diagnosis of Keratitis Lesions Based on Target Localization or Segmentation Algorithms. This method first uses target localization or segmentation algorithms to crop the region of interest (ROI) of the cornea or corneal-conjunctival region from the slit-lamp image. Since the original slit-lamp image contains the entire eye, non-lesion areas such as eyelashes and eyelids, acting as noise, inevitably interfere with the deep learning model, affecting the accuracy of keratitis diagnosis. After automatic localization or segmentation, the ROI related to the keratitis lesion is cropped and input into the CNN network. This allows the CNN to focus on the ROI related to keratitis rather than being distracted by information from the entire image, thereby identifying keratitis pathological features such as corneal opacity, corneal macula, and corneal hyperemia, effectively improving the accuracy and stability of diagnosis.

[0038] Solution 3: Intelligent Keratitis Diagnosis Based on Multiple Image Patch Voting. This method randomly segments the original slit-lamp image into multiple image patches, allowing each patch to focus on a local area of ​​the cornea while avoiding interference from irrelevant areas such as eyelids and eyelashes. Each image patch is individually input into a deep CNN for feature extraction and keratitis classification. By integrating the classification results of each image patch through a majority voting mechanism, the impact of misclassification of a single image patch on the overall diagnosis can be reduced, enhancing the identification ability and robustness of early keratitis lesions and providing a feasible diagnostic scheme for keratitis in clinical practice. Furthermore, this method is not only adaptable to images of different resolutions but also applicable to the automatic diagnosis of other corneal abnormalities besides keratitis.

[0039] Solution 4: Intelligent Keratitis Screening Method Based on Deep Meta-Learning: This method can intelligently diagnose keratitis using very few slit-lamp images. It acquires and labels slit-lamp images, constructs an auxiliary intelligent screening model using a deep learning classification algorithm, inputs the slit-lamp images into the model to extract keratitis features, and trains the model. Combining corneal images captured by the slit-lamp and mobile devices, it uses a meta-learning strategy to extract features from the slit-lamp images, continuously training and updating the model until convergence. This solution is suitable for keratitis diagnosis when slit-lamp image samples are limited.

[0040] However, the aforementioned intelligent keratitis screening technology still faces challenges in practical applications, including reliance on computational resources and risks related to medical privacy data compliance. Specifically, it cannot achieve decentralization to conduct large-scale screenings while ensuring data privacy. Universal keratitis screening often requires primary care physicians to interact with large groups of people during real-time image capture. Cloud-based solutions involve image encoding, long-distance network transmission, server queuing calculations, and result feedback, resulting in significant cumulative latency. This prevents the device from providing real-time "what you see is what you get" assistance in the viewfinder, severely impacting the diagnostic efficiency of large-scale screenings conducted by doctors using handheld devices. Eye images contain rich biometric information, representing highly sensitive personal medical privacy. Under a centralized cloud architecture, raw images must leave the controlled physical terminal and enter the public network environment, posing a significant risk of interception, tampering, or illegal disclosure during data transmission, relay, and cloud storage.

[0041] With increasingly stringent global regulations on medical data protection, localized data processing, where data doesn't leave the device, has become a rigid requirement. Furthermore, centralized solutions heavily rely on stable, high-bandwidth network environments. However, primary healthcare institutions, community screening sites, or remote areas with high rates of keratitis often face unstable network coverage or even complete network outages. Without real-time network support, cloud-based AI systems will be completely paralyzed, unable to provide truly mobile, anytime, anywhere healthcare services. Maintaining high-throughput cloud graphics processing unit (GPU) clusters also requires significant power and maintenance costs. For large-scale group screening tasks, centralized architectures face enormous bandwidth load pressure, and system expansion costs increase exponentially with the number of users.

[0042] To address the aforementioned technical problems, this invention provides a portable real-time keratitis screening device based on lightweight edge computing.

[0043] The following is combined Figures 1 to 6 This invention describes a portable real-time keratitis screening device based on lightweight edge computing.

[0044] Figure 1 This is one of the structural schematic diagrams of the portable real-time keratitis screening device based on lightweight edge computing provided by the present invention, such as... Figure 1 As shown, this portable real-time keratitis screening device based on lightweight edge computing includes: Optical acquisition unit 101, control unit 102 and interaction unit 103; The interaction unit 103 is used to receive image acquisition commands; The optical acquisition unit 101 includes a macro camera module, and the optical acquisition unit 101 is used to call the macro camera module to acquire an eye image of the target object based on the image acquisition command. The control unit 102 includes an edge development board, on which a lightweight image recognition algorithm is deployed on the neural processing unit hardware accelerator. The control unit 102 is used to perform real-time keratitis screening processing on the eye image based on the lightweight image recognition algorithm to obtain the keratitis screening result of the target object.

[0045] Specifically, the interaction unit is used to interact with the user. For example, the interaction unit includes at least one of a voice module, a touch screen module, and a button module. The voice module is used to receive image acquisition commands input by the user's voice, the touch screen module is used to receive image acquisition commands input by the user through touch operation, and the button module is used to receive image acquisition commands input by the user through button operation. Furthermore, the interaction unit transmits the acquired image acquisition commands to the optical acquisition unit through a control unit.

[0046] Specifically, the optical acquisition unit includes a macro camera module for capturing or acquiring images. Upon receiving an image acquisition command, the optical acquisition unit responds by activating the macro camera module to acquire an image of the target object's eyes and transmitting the image to the control unit.

[0047] Specifically, the edge development board can be the Orange Pi 5 Ultra edge-side motherboard. The lightweight image recognition algorithm can be a pre-trained keratitis recognition algorithm that has undergone model lightweighting. Before putting this portable real-time keratitis screening device based on lightweight edge computing into use, the keratitis recognition algorithm can be pre-trained, and then the model of the keratitis recognition algorithm can be lightweighted to obtain a lightweight image recognition algorithm. This lightweight image recognition algorithm can then be deployed on the Neural Processing Unit (NPU) hardware accelerator of the edge development board.

[0048] After obtaining the eye image, the control unit inputs the eye image into a lightweight image recognition algorithm, which processes the eye image to obtain the keratitis screening result of the target object.

[0049] In addition, the keratitis screening results of the target object can also be output through the interactive unit, such as reading out the keratitis screening results through the voice module of the interactive unit, or displaying the keratitis screening results through the touch screen module of the interactive unit.

[0050] For example, see Figure 2 , Figure 2This is the second structural schematic diagram of the portable real-time keratitis screening device based on lightweight edge computing provided by the present invention: The portable real-time keratitis screening device based on lightweight edge computing uses a 3D printed integrated handheld shell as the physical carrier, and the internal components work together through a specific spatial layout and action relationship: (1) Includes Orange Pi 5 Ultra development board and main control board (MAIN PCBA) control unit: The portable real-time keratitis screening device based on lightweight edge computing has an Orange Pi 5 Ultra development board embedded in the back end as the main control unit. It is equipped with an NPU that can process 6 trillion operations per second (6 TOPS) and is responsible for handling complex artificial intelligence inference tasks, namely real-time keratitis screening. The Orange Pi 5 Ultra development board is connected to the main control board. The main control board is connected to the interaction unit.

[0051] (2) Interactive unit: used for real-time image preview, display of diagnostic results and auxiliary interface operation.

[0052] (3) Optical acquisition unit: The portable real-time keratitis screening device based on lightweight edge computing integrates a macro camera module at the front end for image acquisition.

[0053] The portable real-time keratitis screening device based on lightweight edge computing provided by this invention significantly improves inference efficiency by deploying a lightweight image recognition algorithm on the neural processing unit hardware accelerator of the edge development board. By realizing a closed loop from framing and preprocessing to inference result output locally on the device, the response gaps caused by network transmission are completely eliminated, achieving millisecond-level "what you see is what you get" real-time framing assistance, which greatly improves the efficiency of doctors in capturing subtle lesions. Through the lightweight image recognition algorithm, high-performance image feature expression and classification are achieved with a very small number of parameters, so that complex medical image recognition tasks do not need to transmit private data via the Internet, but are completed in the memory of the terminal device and within the neural processing unit. This data processing mechanism that does not leave the device physically blocks the path of interception or tampering, and solves the risk of leakage.

[0054] In some embodiments, the optical acquisition unit 101 further includes a time-of-flight sensor module; The optical acquisition unit 101 is also used to acquire the flight time of the light pulse between the corneal surface of the target object and the macro camera module through the time-of-flight sensor module; The control unit 102 is also used to calculate the distance between the corneal surface and the macro camera module based on the flight time, and generate a capture command when the distance is within the macro imaging range; The optical acquisition unit 101 is also used to respond to the capture command by calling the macro camera module to automatically focus and acquire the eye image.

[0055] Specifically, see Figure 3 , Figure 3 This is a physical schematic diagram of the Time-of-Flight (ToF) sensor module provided by this invention: The ToF sensor module is a key component for achieving high-precision assisted focusing and image quality control, optimizing imaging feedback by monitoring the physical distance between the lens and the eye in real time. See also Figure 2 The time-of-flight sensor module is located near the macro camera module, such as on the surface of the macro camera module, to assist the macro camera module in high-precision focusing.

[0056] Specifically, the macro imaging range is used to determine whether a shot can be taken, and it can be set according to needs, such as 3-5 centimeters.

[0057] In practical applications, the time-of-flight sensor module utilizes time-of-flight detection technology. An internal optical emitting unit emits infrared light pulses of a specific frequency towards the subject's eye (the corneal surface of the target object), and a receiving array captures the reflected signal (the signal from the infrared light pulse reflected to the macro camera module). Based on the time interval between the emitted and reflected signals, the flight time Δt is obtained and fed back to the control unit.

[0058] After obtaining the time-of-flight (Δt), the control unit calculates the (absolute) distance between the front end of the device (macro camera module) and the corneal surface based on Δt and the speed of light. Logically, the time-of-flight sensor module and the macro camera module form a closed-loop linkage. When the control unit detects that the distance is within the preset macro imaging range (e.g., 3-5 cm), the main control board generates a capture command, immediately triggering the macro camera module to automatically focus and compensate, in order to acquire an image of the eye.

[0059] Optionally, the control unit 102 is specifically used to: multiply the flight time, the speed of light, and a preset value to obtain the distance between the corneal surface and the macro camera module. For example, the preset value is 1 / 2.

[0060] After obtaining the flight time Δt, the control unit can calculate the distance d between the corneal surface and the macro camera module using the following formula: d = (c × Δt) / 2; Where c is the speed of light.

[0061] In this embodiment of the invention, a time-of-flight sensor module is set up to determine when image acquisition is performed. This physical ranging method has a faster response speed and stronger anti-interference ability compared with traditional contrast focusing algorithms. Especially in scenarios where corneal lesions are small and have inconspicuous features, it can ensure that the acquired images are always within the optimal recognition clarity range of the lightweight image recognition algorithm, thus guaranteeing the accuracy and stability of diagnosis from the hardware level.

[0062] In some embodiments, the macro camera module includes a macro camera and a supplementary lighting device disposed around the macro camera; The supplemental lighting device is used to adaptively adjust the brightness according to the intensity of ambient light.

[0063] In practical applications, to ensure consistent lighting in the acquisition environment, a dedicated ring light solution, i.e., supplementary lighting equipment, can be configured around the macro camera. For example... Figure 2 A ring of fill lights is set around the outside of the macro camera module.

[0064] In some embodiments, the portable real-time keratitis screening device based on lightweight edge computing further includes: a battery management module; The battery management module includes a battery pack, a charging module for energy management of the battery pack, and a step-down converter module. The battery management module is used to convert the voltage output by the battery pack into a stable DC voltage through the step-down conversion module, and to power the optical acquisition unit 101, the control unit 102 and the interaction unit 103 through the DC voltage.

[0065] Specifically, the battery management module can power an 18650 (lithium) battery pack system, and the battery pack can be an 18650 (lithium) battery pack. The charging module can be a TP4056 charging module. The buck converter module can be a DC-DC (Direct Current to Direct Current) buck converter module.

[0066] In practical applications, see Figure 2 The battery management module has a built-in 18650 battery pack as its energy source and manages energy through the TP4056 charging module. In other words, the power management system (battery management module) integrates the battery pack, charging module and step-down conversion module to build a stable local energy supply system.

[0067] Specifically, the battery output is connected to a DC-DC step-down converter module based on the MP1584EN chip, which converts the unstable battery voltage (the voltage output by the battery pack) into a stable 5V / 3.3V DC power (stable DC voltage) to power the motherboard and various sensors, namely the optical acquisition unit, control unit and interaction unit.

[0068] In this embodiment of the invention, power is supplied through a battery management module, which ensures the battery life of the portable real-time keratitis screening device based on lightweight edge computing. In addition, the high-efficiency DC-DC buck scheme based on MP1584EN further extends the battery life of the handheld device.

[0069] In some of these embodiments, the buck converter module includes a synchronous buck converter, a transient voltage suppressor, a multi-stage decoupling filter capacitor, an energy storage inductor, and an output filter capacitor bank. The transient voltage suppressor and the multi-stage decoupling filter capacitor are respectively connected to the input terminal of the synchronous buck converter; The energy storage inductor is connected to the output terminal of the synchronous buck converter; The current corresponding to the voltage output by the battery pack flows into the synchronous buck converter for buck conversion through the transient voltage suppressor and the multi-stage decoupling filter capacitor to obtain the DC voltage. When the internal high-side power transistor of the synchronous buck converter is turned on, the current corresponding to the DC voltage flows into the optical acquisition unit 101, the control unit 102 and the interaction unit 103 through the energy storage inductor.

[0070] Specifically, the synchronous buck converter can be the TPS6216x synchronous buck converter.

[0071] See Figure 4 , Figure 4 This is a circuit diagram of the buck converter module provided by the present invention: The buck converter module integrates a TPS6216x synchronous buck converter (U1) to construct a stable 3.3V (V) digital logic power supply loop. The input terminals of U1 (pins 1 to 4) are equipped with a transient voltage suppressor VR1 (CG0603MLC-05E) and multi-stage decoupling filter capacitors (C1 and C2, where C1 is 22 microfarads (μF) and C2 is 0.1μF), effectively suppressing the impact of input power supply ripple and transient surges on sensitive downstream components (mainly the main control board and development board of the control unit). The voltage provided by the battery pack is transmitted to the buck converter module through the 5VCC interface.

[0072] When the high-side power transistor inside U1 is turned on, current flows through the energy storage inductor L1 (3.3 μH) to supply power to the load via the PG_3.3VCC port, establishing a magnetic field in the inductor to store energy. Simultaneously, it charges the output filter capacitor bank (C3, C4, and C5, where C3 and C4 are both 47 μF, and C5 is 0.01 μF). When the high-side power transistor inside U1 is turned off, U1 opens the freewheeling circuit through the internally integrated synchronous rectifier diode, releasing energy by utilizing the non-abrupt current characteristic of the inductor, ensuring continuous output current and extremely low voltage ripple. The buck converter module uses the VOS pin of U1 to achieve closed-loop feedback of the output voltage, dynamically adjusting the internal switching duty cycle.

[0073] It should be noted that the TPS6216x synchronous buck converter has 9 pins, and the functions of each pin are as follows: Pin 1 (Power Ground, PGND) is the path for high current to flow back to the power supply and is usually connected to the main ground plane of the circuit board.

[0074] Pin 2 (Input Voltage, VIN), the power input, is used to receive the raw voltage from 5VCC to power the internal power switches of the synchronous buck converter.

[0075] Pin 3 (Enable, EN) is the enable / disable control pin. A high level on pin 3 enables the synchronous buck converter, while a low level disables it. Figure 4 Connecting it to VIN means that the system starts working as soon as it is powered on.

[0076] Pin 4 (Analog Ground, AGND) is the reference point for the internal control logic of the synchronous buck converter. To reduce interference, it is usually connected to PGND at a single point.

[0077] Pin 5 (Feedback, FB), the feedback pin, is used in fixed output versions (such as this one). Figure 4 In the 3.3V version, this pin is usually directly grounded or used as an internal compensation reference.

[0078] Pin 6 (Output Voltage Sense, VOS) is used for output voltage sampling. It is directly connected to the 3.3V output point (3.3VCC) at the back end of L1 and is used to monitor the output voltage in real time to maintain stability.

[0079] Pin 7 (Switch, SW) is the switching node. It connects the internal power transistor and the external inductor L1 and is the core node for high-speed voltage switching.

[0080] Pin 8 (Power Good, PG) is an open-drain output pin that goes high when the output voltage reaches the rated value (with pull-up resistor R1), and outputs a high level to indicate that the power supply is stable.

[0081] Pin 9 (Thermal Pad, T-GND), also known as the heat dissipation pad, is located at the bottom of the synchronous buck converter. In addition to electrical grounding, it is mainly used to conduct heat from inside the synchronous buck converter to the copper foil of the printed circuit board (PCB) to prevent overheating.

[0082] Furthermore, the TPS6216x buck converter (U1) achieves efficient voltage conversion through its internally integrated complementary high-side and low-side power switches. During the energy storage phase, the control circuit drives the high-side power switch to turn on and the low-side power switch to turn off, allowing the 5V power from the input to flow into the energy storage inductor L1 via the switching node. The inductor current simultaneously supplies power to the load and charges the output capacitor while establishing a magnetic field to store energy. During the energy release phase, i.e., synchronous rectification, the high-side power switch turns off and the low-side power switch turns on synchronously to form a freewheeling loop. Utilizing the non-abrupt physical characteristic of the inductor current, magnetic energy is continuously converted into electrical energy and released to the load, effectively replacing the external freewheeling diode in traditional buck circuits and significantly reducing conduction losses. Finally, the output voltage is sampled in real time via the VOS pin, and the switching duty cycle of the power switches is dynamically adjusted according to a preset value, achieving precise energy distribution at megahertz (MHz) high-frequency switching. This ensures continuous output current while providing high-efficiency, low-ripple, and stable 3.3V power support to the back-end core processor.

[0083] In this embodiment of the invention, a high-efficiency and highly integrated step-down strategy is formed by using a synchronous step-down converter, a transient voltage suppressor, a multi-stage decoupling filter capacitor, an energy storage inductor, and an output filter capacitor bank. This ensures a highly stable output voltage under complex load conditions, provides clean and reliable power support for the edge development board's main control core and related high-speed digital peripherals, and guarantees the stability of the device under high load operation.

[0084] In some embodiments, the buck converter module further includes a pull-up resistor and an indicator circuit; The pull-up resistor is connected to the output terminal of the synchronous buck converter and is used to output the status information of the synchronous buck converter; The indicator circuit is used to display the status information of the synchronous buck converter through a two-stage LED.

[0085] In practical applications, see Figure 4The status signal, i.e. the status of synchronous buck converter U1, is output through the PG (Power Good) pin in conjunction with the pull-up resistor R1 (22 kΩ). The indicator circuit consisting of LED1 (Light-Emitting Diode, LED) and resistor R2 (220 Ω) is driven through the 3.3VCC port to realize real-time visual monitoring of the power supply health status (synchronous buck converter U1).

[0086] In some embodiments, the lightweight image recognition algorithm is a lightweight version of the MobileNetV4 algorithm; The lightweight MobileNetV4 algorithm is the RKNN format MobileNetV4 algorithm; The RKNN format MobileNetV4 algorithm is obtained by deeply optimizing the ONNX format MobileNetV4 algorithm using the RKNN-Toolkit toolchain; The ONNX format MobileNetV4 algorithm is obtained by converting the PTH format MobileNetV4 algorithm trained with PyTorch.

[0087] Specifically, the MobileNetV4 (MobileNet4) image recognition algorithm. The RKNN (Rockchip Neural Network) format is a neural network model inference format specifically designed for artificial intelligence chips, used to efficiently run deep learning models on the NPU, especially suitable for edge computing and embedded devices.

[0088] In practical applications, see Figure 5 , Figure 5This is a flowchart illustrating the deployment of the MobileNetV4 algorithm provided in this embodiment of the invention: The MobileNetV4 algorithm is trained using PyTorch (i.e., PyTorch training) to obtain a trained MobileNetV4 algorithm, namely, a MobileNetV4 algorithm in PTH format (.pth). Since PyTorch uses a dynamic graph mechanism, while edge inference engines typically require static graphs, the first step in the conversion is to convert the PTH format MobileNetV4 algorithm into an ONNX (.onnx) format MobileNetV4 algorithm through the ONNX (Open Neural Network Exchange) export process, i.e., converting the model to ONNX; then, the RKNN-Toolkit toolchain is used to perform deep optimization on the ONNX format MobileNetV4 algorithm, finally generating an RKNN (.rknn) format MobileNetV4 algorithm (i.e., converting the model to RKNN through the RKNN export process), and then deploying it on an Orange Pi 5 Ultra for asynchronous inference.

[0089] To deploy the server-trained MobileNetV4 algorithm on a development board for low-power real-time inference, a format conversion from .pth to .onnx to .rknn is required, which can reduce the model size by more than 90%. Converting to the ONNX intermediate format fixes the MobileNetV4 algorithm's topology, eliminates dynamic branches, facilitates subsequent operator fusion, and ensures that PyTorch operators (such as convolution and activation functions) match the ONNX standard operator library. This prepares the RKNN toolchain for parsing, which can fuse adjacent operators (such as Conv, BatchNorm, and ReLU) into a single hardware instruction. This reduces the number of data transfers between memory and computing units, increases the ratio of arithmetic operations to memory accesses, and leverages the NPU's superior support for fixed-point operations (INT8) compared to floating-point operations (FP32). Mapping floating-point numbers to the [-128, 127] range greatly alleviates memory bandwidth limitations and improves real-time inference efficiency.

[0090] In some embodiments, the control unit 102 is specifically used for: The eye image is cropped and standardized to obtain a corneal image; The corneal image is input into the lightweight image recognition algorithm, and the features of the corneal image are extracted by a general inverted residual block. The feature extraction results are analyzed to obtain the keratitis screening results, which are classification probability matrices for three categories: keratitis, normal, and other corneal abnormalities.

[0091] In practical applications, after obtaining the eye image, it needs to be cropped and standardized to obtain the corneal image: the 3D color image acquired by the optical acquisition unit is cropped to a size of 3x448x448 and input into the lightweight image recognition algorithm. Therefore, the image dimension input to the lightweight image recognition algorithm is 3x448x448. First, it is passed through a 3x3 convolution with a stride of 2 to reduce the image size to 224x224, and then passed through four Universal Inverted Residual Blocks with a stride of 2. Bottleneck (UIB) is used to further compress the image size to 14x14. A 1x1 convolution is then used to increase the feature map dimension to 1024, condensing and fusing the previously learned scattered features. At this point, the image dimension in the model is 14x14x1024. After a global average pooling operation, the image dimension becomes 1x1x1024. Finally, a fully connected layer maps all feature maps to a 1x1x3 dimension, corresponding to the three categories: Keratitis, Other corneal abnormalities, and Normal.

[0092] The general inverse residual block mainly consists of four core components: ExtraDW, located at the front end, is used to extract spatial features before changing the number of channels; 1x1 expanded convolution is used to enlarge the number of input channels C_{in} to a higher dimension (usually 4xC_{in}), increasing the feature representation space; MiddleDW is used to perform 3×3 or 5×5 depth convolution in the expanded high-dimensional space, responsible for the main spatial modeling; and 1x1 projected convolution is used to compress high-dimensional features back to the number of output channels.

[0093] When a feature map enters the UIB (User Interface Block), the following operations are performed: First, local features are extracted using depthwise convolution for spatial pre-filtering. Then, the channel dimension is increased using a 1x1 dilated convolution. Next, a sliding window is calculated for each channel dimension to model the depthwise convolution. Finally, the dimensions are restored using 1x1 projection. The learned multi-channel features are then linearly fused to obtain the output. If the input and output dimensions are the same, the original input and the processed output are added together to obtain the final output.

[0094] In this embodiment of the invention, a lightweight image recognition algorithm combines Depthically Separable Convolution (DW-Conv) with Optimized Neural Architecture Search (NAS), achieving high-performance image feature representation and classification with a minimal number of parameters. This allows complex medical image recognition tasks to be completed entirely within the terminal device's memory and NPU without streaming private data over the internet. This "data not leaving the device" processing mechanism physically blocks the path of interception or tampering, complying with the most stringent global medical data security compliance standards.

[0095] In some embodiments, the interaction unit 103 includes a touch display screen.

[0096] Specifically, the touch module can be a touch display screen for real-time image preview, display of diagnostic results, and auxiliary interface operation.

[0097] Optionally, the touch display screen is used to generate the image acquisition command in response to a user's touch operation.

[0098] In practical applications, operators (users) can trigger detection commands via touch or physical buttons (such as the screen's Enter key / the product's physical shutter button).

[0099] Optionally, the touch display screen is used to display the eye image and / or the keratitis screening results.

[0100] In practical applications, after capturing an image of the eye, a feedback message can be sent to the interactive interface of the touch screen, which displays the eye image to inform the user that the image has been captured.

[0101] In addition, the keratitis screening results can be displayed via an interactive touchscreen interface.

[0102] like Figure 6 As shown, Figure 6 This is a schematic diagram of the interactive interface provided by the present invention: The left side of the interactive interface displays eye images (i.e., real-time camera images) collected in real time by the portable real-time keratitis screening device, and displays the screening result (Result): Keratitis (87.7%) on the eye images; the right side of the interactive interface displays the operation interface, including the display interface of the test results (waiting for operation, indicating that the current operation has been completed and waiting to detect the next eye image), the capture detection (Enter) control, the exit program (Q) control, and the operation feedback area. In addition, a reminder message is also displayed at the bottom of the operation interface, such as "Tip: Press Enter to start the detection, Q to exit. Mouse clicks and keyboard inputs will be displayed here as feedback."

[0103] In this way, users or target objects can obtain the current portable real-time keratitis screening results in a timely and accurate manner through the touch screen, thereby improving user satisfaction.

[0104] In some embodiments, the control unit 102 is connected to the interaction unit 103 via a mobile industrial processor interface.

[0105] Specifically, the touch display can be a Mobile Industry Processor Interface (MIPI) touch display.

[0106] See Figure 2 The main control board in the control unit is connected to the MIPI touch screen via the MIPI interface for real-time image preview, display of diagnostic results, and auxiliary interface operation.

[0107] The following is combined Figure 2 The present invention will further describe the portable real-time keratitis screening device based on lightweight edge computing provided by the present invention.

[0108] Figure 2 The core of the portable real-time keratitis screening device based on lightweight edge computing is based on the MobileNetV4 architecture. High-performance real-time inference is achieved through customized design to address hardware bottlenecks. The lightweight MobileNet V4 image recognition algorithm is deployed on the edge development board of Orange Pi 5 Ultra after model lightweighting for the RK3588 chip. The device uses a front-facing camera with an LED aperture to capture clear eye images as input to the model for inference to obtain the most likely keratitis screening category. All operations are visualized on the MIPI touch screen on the back of the device, which provides real-time display of image acquisition and inference results. Combined with a power management module, it can be used as a handheld mobile device for long-term eye screening.

[0109] Specifically, system initialization and real-time preview: After the lightweight edge computing portable real-time keratitis screening device is powered on, the Orange Pi5 starts the device (enters use mode) and displays the interactive interface on the MIPI touch screen. The macro camera starts, and the ring light adaptively adjusts its brightness according to the ambient light intensity, displaying the high-definition image from the camera in real time on the screen.

[0110] Intelligent capture and preprocessing: The operator triggers the detection command via touch or physical buttons (Enter key on the screen / physical shooting button on the product). Next, ToF ranging information is used to determine whether the acquisition distance is within the macro focal length range (3-5cm). If the condition is met, the image is captured and cropped and standardized. After the capture is completed, the macro intelligent capture mode is ended again via touch or physical buttons.

[0111] Edge-side NPU inference: The preprocessed image is input into the MobileNetV4 model loaded with the NPU. The model extracts features through the UIB structure and outputs classification probability matrices for the three categories: Keratitis, Normal, and Other.

[0112] Results Feedback and Interaction: Detection results and their corresponding confidence levels (e.g., 42.0% Normal) are displayed and updated in real-time on the interactive interface. The entire inference process is completed locally in milliseconds, and diagnostic data is encrypted and stored locally without needing to be uploaded to the cloud.

[0113] This invention provides a portable real-time keratitis screening device based on lightweight edge computing. Addressing the significant latency issues in cloud transmission and computation inherent in existing solutions, it deploys the lightweight MobileNetV4 algorithm on the NPU hardware accelerator of the Orange Pi 5Ultra edge-side motherboard. Technically, MobileNetV4 introduces a general inverse bottleneck structure and a Mobile MQA attention module, exhibiting Pareto-optimal performance on mobile computing platforms and significantly improving inference efficiency. By implementing a closed loop locally on the device from framing and preprocessing to inference result output, it completely eliminates response gaps caused by network transmission, achieving millisecond-level "what you see is what you get" real-time framing assistance. This greatly improves the efficiency of doctors in capturing subtle lesions, achieving an ultimate real-time interactive innovation between MobileNetV4 and edge acceleration.

[0114] This invention provides a portable real-time keratitis screening device based on lightweight edge computing. To mitigate the risk of data leakage during public network transmission and cloud storage of raw images in centralized architectures, it achieves a fully localized data loop. Technically, MobileNetV4 combines depthwise separable convolution (DW-Conv) with optimized neural network architecture search (NAS), achieving high-performance image feature representation and classification with minimal parameters. This allows complex medical image recognition tasks to be completed entirely within the terminal device's memory and NPU without streaming private data over the internet. This "data not leaving the device" processing mechanism physically blocks the path of interception or tampering, complying with the world's most stringent medical data security compliance standards, thus achieving a physically isolated, edge-side medical privacy protection technology innovation.

[0115] This invention provides a portable real-time keratitis screening device based on lightweight edge computing. Addressing the pain point of unstable network coverage in primary healthcare institutions leading to the susceptibility of AI systems to failure, this device constructs a completely decentralized, independent hardware system. Technically, this solution directly burns the quantized lightweight model to the edge-side main controller, giving the device complete "offline survivability." Combined with an integrated 18650 lithium battery power supply system, integrated 3D-printed medical packaging, and a local visual user interface, the development board's NPU maintains full AI diagnostic performance even in environments without network access, offline, or extremely remote locations. This innovation eliminates the rigid dependence on high-bandwidth infrastructure, ensuring true mobility and anytime, anywhere medical services in any geographical environment—a robust offline "all-scenario screening" architecture innovation.

[0116] As can be seen, the portable real-time keratitis screening device based on lightweight edge computing proposed in this invention has the following significant advantages compared with traditional centralized cloud diagnostic solutions and existing portable devices: Pareto optimal real-time detection performance: The MobileNetV4 used in this invention benefits from the optimized MobileMQA attention module, resulting in an inference speed improvement of approximately 39%.

[0117] Extremely high diagnostic accuracy: By utilizing MobileNetV4's general inverse bottleneck block, MobileNetV4 can more effectively balance the mixing of spatial and channel features, thereby more accurately extracting small and irregular lesion features such as corneal opacities and corneal lesions, enabling auxiliary diagnostic accuracy comparable to complex models on the server side on edge devices.

[0118] Ultimate medical privacy protection: Because MobileNetV4 enables fully localized inference, patients' eye images do not need to be streamed to the cloud via the public internet, thus avoiding the risk of leakage of sensitive medical data at the physical source.

[0119] Exceptional portability and all-scenario adaptability: By integrating the Orange Pi 5 Ultra edge controller, 18650 battery pack, and efficient power management module, it can operate stably in grassroots and remote areas without network coverage. Combined with real-time physical ranging via a ToF sensor and macro focusing linkage, it effectively solves the problem of traditional focusing algorithms failing due to the minute size of corneal lesions, ensuring the quality of the image input by the algorithm.

[0120] Low power consumption and high cost-effectiveness: By converting the model from FP32 to INT8 format using RKNN quantization technology, the memory bandwidth requirements and computing power consumption are greatly reduced. Combined with the MP1584EN's high-efficiency DC-DC buck solution, the battery life of handheld devices is further extended. All inference is performed on edge devices, eliminating the expensive cloud server leasing and high bandwidth maintenance costs, which is conducive to large-scale deployment and use.

[0121] This invention also provides a portable real-time keratitis screening method based on lightweight edge computing. This method is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. The portable real-time keratitis screening method based on lightweight edge computing provided by this invention will be described below. The portable real-time keratitis screening method based on lightweight edge computing described below can be referred to in correspondence with the portable real-time keratitis screening device based on lightweight edge computing described above.

[0122] Figure 7 This is a schematic flowchart of the portable real-time keratitis screening method based on lightweight edge computing provided by the present invention. Figure 7 As shown, this portable real-time keratitis screening method based on lightweight edge computing is applied to a portable real-time keratitis screening device based on lightweight edge computing. The portable real-time keratitis screening device based on lightweight edge computing includes an optical acquisition unit, a control unit, and an interaction unit. The optical acquisition unit includes a macro camera module, and the control unit includes an edge development board. A lightweight image recognition algorithm is deployed on the neural processing unit hardware accelerator of the edge development board. The method includes: Step 701: Receive image acquisition command through the interaction unit; Step 702: Using the optical acquisition unit, based on the image acquisition command, the macro camera module is invoked to acquire an image of the target object's eye; Step 703: Using the control unit, based on the lightweight image recognition algorithm, perform real-time keratitis screening on the eye image to obtain the keratitis screening result of the target object.

[0123] The portable real-time keratitis screening method based on lightweight edge computing provided by this invention significantly improves inference efficiency by deploying a lightweight image recognition algorithm on the neural processing unit hardware accelerator of the edge development board. By realizing a closed loop from framing and preprocessing to inference result output locally on the device, the response gaps caused by network transmission are completely eliminated, achieving millisecond-level "what you see is what you get" real-time framing assistance, which greatly improves the efficiency of doctors in capturing subtle lesions. The lightweight image recognition algorithm achieves high-performance image feature expression and classification with a very small number of parameters, so that complex medical image recognition tasks do not need to transmit private data via the Internet, but complete the entire cycle of calculation within the terminal device's memory and neural processing unit. This data processing mechanism that does not leave the device physically blocks the path of interception or tampering, solving the risk of leakage.

[0124] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 8 As shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830, and a communications bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other through the communications bus 840. The processor 810 can call logical instructions in the memory 830 to execute a portable real-time keratitis screening method based on lightweight edge computing. This method is applied to a portable real-time keratitis screening device based on lightweight edge computing. The device includes an optical acquisition unit, a control unit, and an interaction unit. The optical acquisition unit includes a macro camera module, and the control unit includes an edge development board. A lightweight image recognition algorithm is deployed on the neural processing unit hardware accelerator of the edge development board. The method includes: receiving an image acquisition command through the interaction unit; using the optical acquisition unit to call the macro camera module to acquire an eye image of a target object based on the image acquisition command; and using the control unit to perform real-time keratitis screening processing on the eye image based on the lightweight image recognition algorithm to obtain the keratitis screening result of the target object.

[0125] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a portable peripheral readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a portable peripheral device to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0126] The present invention also provides an electronic device including a memory and a processor, the memory storing a computer program and the processor being configured to run the computer program to perform the steps in any of the above method embodiments.

[0127] On the other hand, in conjunction with the portable real-time keratitis screening method based on lightweight edge computing provided in the above embodiments, the present invention also provides a non-transitory computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in any of the above embodiments of the portable real-time keratitis screening method based on lightweight edge computing.

[0128] In another aspect, in conjunction with the portable real-time keratitis screening method based on lightweight edge computing provided in the above embodiments, the present invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0129] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0130] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a portable peripheral device to execute the methods described in the various embodiments or some parts of the embodiments.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A portable real-time keratitis screening device based on lightweight edge computing, characterized in that, include: Optical acquisition unit, control unit, and interaction unit; The interactive unit is used to receive image acquisition commands; The optical acquisition unit includes a macro camera module, and the optical acquisition unit is used to call the macro camera module to acquire an image of the target object's eye based on the image acquisition command; The control unit includes an edge development board, on which a lightweight image recognition algorithm is deployed on the neural processing unit hardware accelerator. The control unit is used to perform real-time keratitis screening processing on the eye image based on the lightweight image recognition algorithm to obtain the keratitis screening result of the target object.

2. The portable real-time keratitis screening device based on lightweight edge computing according to claim 1, characterized in that, The optical acquisition unit also includes a time-of-flight sensor module; The optical acquisition unit is also used to acquire the flight time of the light pulse between the corneal surface of the target object and the macro camera module through the time-of-flight sensor module; The control unit is also used to calculate the distance between the corneal surface and the macro camera module based on the flight time, and generate a capture command when the distance is within the macro imaging range; The optical acquisition unit is also used to respond to the capture command by calling the macro camera module to automatically focus and acquire the eye image.

3. The portable real-time keratitis screening device based on lightweight edge computing according to claim 1, characterized in that, Also includes: Battery management module; The battery management module includes a battery pack, a charging module for energy management of the battery pack, and a step-down converter module. The battery management module is used to convert the voltage output by the battery pack into a stable DC voltage through the step-down converter module, and to power the optical acquisition unit, the control unit and the interaction unit through the DC voltage.

4. The portable real-time keratitis screening device based on lightweight edge computing according to claim 3, characterized in that, The buck converter module includes a synchronous buck converter, a transient voltage suppressor, a multi-stage decoupling filter capacitor, an energy storage inductor, and an output filter capacitor bank. The transient voltage suppressor and the multi-stage decoupling filter capacitor are respectively connected to the input terminal of the synchronous buck converter; The energy storage inductor is connected to the output terminal of the synchronous buck converter; The current corresponding to the voltage output by the battery pack flows into the synchronous buck converter for buck conversion through the transient voltage suppressor and the multi-stage decoupling filter capacitor to obtain the DC voltage. When the internal high-side power transistor of the synchronous buck converter is turned on, the current corresponding to the DC voltage flows into the optical acquisition unit, the control unit, and the interaction unit through the energy storage inductor.

5. The portable real-time keratitis screening device based on lightweight edge computing according to claim 4, characterized in that, The step-down converter module also includes a pull-up resistor and an indicator circuit; The pull-up resistor is connected to the output terminal of the synchronous buck converter and is used to output the status information of the synchronous buck converter; The indicator circuit is used to display the status information of the synchronous buck converter through a two-stage LED.

6. The portable real-time keratitis screening device based on lightweight edge computing according to claim 1, characterized in that, The lightweight image recognition algorithm is a lightweight version of the MobileNetV4 fourth-generation image recognition algorithm for mobile networks. The lightweight MobileNetV4 algorithm is the RKNN format MobileNetV4 algorithm; The RKNN format MobileNetV4 algorithm is obtained by deeply optimizing the ONNX format MobileNetV4 algorithm using the RKNN-Toolkit toolchain; The ONNX format MobileNetV4 algorithm is obtained by converting the PTH format MobileNetV4 algorithm trained with PyTorch.

7. The portable real-time keratitis screening device based on lightweight edge computing according to claim 1, characterized in that, The control unit is specifically used for: The eye image is cropped and standardized to obtain a corneal image; The corneal image is input into the lightweight image recognition algorithm, and the features of the corneal image are extracted by a general inverted residual block. The feature extraction results are analyzed to obtain the keratitis screening results, which are classification probability matrices for three categories: keratitis, normal, and other corneal abnormalities.

8. The portable real-time keratitis screening device based on lightweight edge computing according to claim 1, characterized in that, The interactive unit includes a touch screen display. The touch display screen is used to generate the image acquisition command in response to the user's touch operation; Or, the touch display screen is used to display the eye image and / or the keratitis screening result.

9. The portable real-time keratitis screening device based on lightweight edge computing according to claim 2, characterized in that, The control unit is specifically used for: Multiplying the flight time, speed of light, and preset values ​​yields the distance between the corneal surface and the macro camera module.

10. The portable real-time keratitis screening device based on lightweight edge computing according to any one of claims 1-9, characterized in that, The macro camera module includes a macro camera and a supplementary lighting device disposed around the macro camera; The supplemental lighting device is used to adaptively adjust the brightness according to the intensity of ambient light; The control unit is connected to the interaction unit via a mobile industrial processor interface.

Images