Smartphone and deep learning-based test strip cartridge detection system and detection method

Abstract

The application discloses a test paper box detection system and method based on a smart phone and deep learning, and belongs to the mobile terminal image recognition, biological detection auxiliary judgment and artificial intelligence application field. The application provides a two-stage neural network decoupling detection scheme. The system comprises a smart phone end and a server end, the smart phone end is used for collecting and uploading test paper box images, the server end carries out image effectiveness and quality screening through image preprocessing and a quality evaluation module loaded with a first neural network, only the effective images are called to carry out category determination through a detection reasoning module loaded with a second neural network, and finally the detection result is returned to the smart phone end for display. The application realizes automatic identification and interception of abnormal images such as fuzzy images and target-free images, improves the accuracy and robustness of detection, and is suitable for medical treatment, food safety and environmental detection and the like.

Description

Technical Field

[0001] This invention belongs to the fields of mobile terminal image recognition, biological detection auxiliary judgment and artificial intelligence application, specifically involving a test kit detection system and detection method based on smartphones and deep learning. Background Technology

[0002] Test strips, as a rapid and low-cost testing method, are widely used in medical testing, food safety testing, and environmental testing. Current test strip methods typically rely on visual observation of color changes in the color-developing area of ​​the test strip to determine whether the test result is negative, positive, or neutral.

[0003] However, existing technologies have at least the following drawbacks: (1) High subjectivity and large error: Manual visual judgment is highly dependent on the user's experience. Different people have significant differences in their understanding of color depth and color boundary. Especially in weak positive or critical conditions, the misjudgment rate is high. (2) Significantly affected by the shooting environment: factors such as lighting conditions, shooting angle, background color and mobile phone model differences can all interfere with the color development effect of the test strip, resulting in poor stability of the test results; (3) Lack of effective processing mechanism for abnormal images: Some existing test strip detection methods based on image recognition usually assume that the input image is valid. It is difficult to effectively distinguish images that are blurry, misaligned with the test strip box, or do not contain a valid color development area, and it is easy to output incorrect detection results. (4) The test results lack a unified and standardized intelligent judgment process, making it difficult to realize the digital management and remote application of the test results.

[0004] Therefore, there is an urgent need for a detection system and method that can adapt to complex shooting environments, has the ability to recognize abnormal images, and can make stable and intelligent judgments on the color development status of test strip boxes. Summary of the Invention

[0005] The technical problem to be solved by this invention is: how to eliminate the influence of differences in shooting environment on image quality while ensuring high accuracy of test strip detection results, achieve effective identification and interception of abnormal images, and reduce the occupation of invalid images on system computing resources.

[0006] To address the aforementioned problems, this invention provides a test kit detection system and method based on smartphones and deep learning. The specific technical solution is as follows: This invention provides a test kit detection system based on smartphones and deep learning, comprising: The smartphone app is used to capture images of the test kit, upload the raw images to the server, and receive and display the test results. A server-side component, communicatively connected to the smartphone, includes: The image uploading and communication module is used to receive the original image and image identification information uploaded by the smartphone. The image preprocessing module is used to preprocess the received raw image to obtain the image to be evaluated; The image quality assessment module is equipped with a first deep learning neural network, which is used to extract features from the image to be assessed and output the image validity and quality assessment results, including whether the image is valid, blurry, or has no detected target. The detection and reasoning module is equipped with a second deep learning neural network, which is used to perform detection and reasoning on the image to be evaluated when the image validity and quality evaluation result is that the image is valid, and output the category detection result of the test kit; The result determination and feedback module is used to generate a final detection status based on the image validity and quality assessment results or the category detection results, and return the final detection status to the smartphone.

[0007] Optionally, the image preprocessing module performs one or more of the following preprocessing: brightness normalization, contrast normalization, white balance normalization, and resolution unification.

[0008] Optionally, the model size of the first deep learning neural network is smaller than that of the second deep learning neural network.

[0009] This invention also provides a test kit detection method based on smartphones and deep learning, comprising the following steps: Step S1: The smartphone captures the original image of the test kit and uploads the original image and image identification information to the server. Step S2: The server performs integrity verification on the received original image. If data is missing or abnormal, a retake prompt is sent to the smartphone. Step S3: The server preprocesses the original image that has passed the integrity check to generate the image to be evaluated; Step S4: Call the first deep learning neural network to extract features from the image to be evaluated, and output the image validity and quality evaluation results, including whether the image is valid, blurry, or has no detected target; Step S5: Determine whether the image validity and quality assessment result indicates that the image is valid; Step S6: If yes, then call the second deep learning neural network to perform detection inference on the image to be evaluated and output the category detection result of the test kit; Step S7: Generate a final detection status based on the image validity and quality assessment results or the category detection results, and return the final detection status to the smartphone for display.

[0010] Optionally, in step S4, "no target detected" means that no test kit structure or color development area was detected; "blurred image" means that the test kit was detected but the clarity did not meet the confidence threshold.

[0011] Optionally, the second deep learning neural network adopts the ConvNeXt network structure.

[0012] Optionally, the final detection status generated in step S7 includes: image invalid, image blurry, neutral, positive, or negative.

[0013] Optionally, the image failure is the judgment result when no target is detected in step S4; the image blurriness is the judgment result when the image is blurred in step S4.

[0014] Optionally, the neutral result includes the judgment result when the confidence of the second deep learning neural network for both negative and positive judgments is lower than the confidence threshold.

[0015] Optionally, in step S6, only the valid results of the images output in step S4 are qualitatively determined.

[0016] The present invention has the following beneficial effects: First, the present invention performs brightness normalization, contrast normalization, white balance normalization, and resolution unification on the original image through an image preprocessing module. This effectively eliminates the interference of factors such as lighting conditions, shooting angle, background color, and differences in mobile phone models on image quality, significantly improves the system's adaptability to complex shooting environments, and ensures the consistency and accuracy of subsequent image analysis.

[0017] Secondly, this invention utilizes an image quality assessment module equipped with a first deep learning neural network to evaluate the validity and quality of preprocessed images. It accurately outputs three states: valid image, blurred image, or no detection target. This enables intelligent identification and dynamic interception of abnormal images such as blurred images, misaligned test kits, or images lacking valid color development areas. The system only calls the subsequent detection inference module for images with valid evaluation results, thus avoiding erroneous detections caused by abnormal images at the source and significantly improving the robustness and reliability of the detection system. This pre-screening mechanism addresses the shortcomings of existing technologies that assume all input images are valid and cannot process abnormal images, while also effectively reducing the consumption of server computing resources by invalid images.

[0018] Third, this invention employs a two-level neural network architecture that decouples image quality assessment from detection inference, where the model size of the first deep learning neural network is smaller than that of the second deep learning neural network. This architecture organically combines rapid screening with accurate classification: a lightweight front-end network is responsible for quickly intercepting invalid images, while the main network focuses on the fine recognition of the color features of the test kit. Compared to traditional coupled integration schemes, this invention significantly reduces inference latency and the total number of model parameters while ensuring high accuracy of detection results, balancing detection efficiency and system resource consumption, and achieving a balance between high accuracy and low resource consumption.

[0019] Fourth, the detection inference module of this invention is equipped with a second deep learning neural network, which outputs the negative, positive, or neutral category detection results of the test kit. The final detection status, including image failure, image blurriness, neutral, positive, or negative, is generated by the result judgment and feedback module based on the quality assessment results or detection inference results. The neutral result includes the judgment result when the confidence level of the second deep learning neural network for both negative and positive assessments is below the confidence threshold. The confidence level verification and security degradation mechanism avoids the risk of false positives or false negatives caused by the deep learning model forcibly classifying under edge features, further ensuring the reliability of the detection results. The final detection status is displayed intuitively on a smartphone, and the user is prompted to retake the photo if the image is blurry or failed, forming a closed-loop interactive experience and ensuring a smooth detection process.

[0020] Fifth, the entire system of this invention is based on a smartphone app to achieve image acquisition and result display, with the server performing the core calculations. Users do not need to purchase dedicated equipment, resulting in low deployment costs, simple operation, and suitability for large-scale promotion and application. At the same time, the detection results are output in digital form, which facilitates subsequent data management and remote application. Attached Figure Description

[0021] 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 The overall flowchart of the detection method provided by the present invention is shown.

[0023] Figure 2 The diagram shows the ConvNeXt network architecture of the detection system provided by this invention.

[0024] Figure 3 This is a schematic diagram of the smartphone detection interface provided by the present invention. Detailed Implementation

[0025] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1 This embodiment provides a test kit detection system based on smartphones and deep learning. The system includes a smartphone terminal and a server terminal, which are connected via a wireless network.

[0027] The smartphone app is used to capture images of the test kit, upload the raw images to the server, and receive and display the test results returned by the server. Specifically, a mini-program runs on the smartphone app. Users access the testing function through the mini-program interface, and guided by the interface, place the test kit to be tested within the preset shooting area. Then, the smartphone's built-in camera captures a raw image containing the test kit and its color development area. After acquisition, the smartphone uploads the raw image along with image identification information (such as user ID, device model, acquisition time, etc.) to the server. The mini-program-guided shooting standardizes user operation, reduces image quality differences caused by improper shooting angles and distances, and lays the foundation for accurate subsequent testing.

[0028] The server side includes an image uploading and communication module, an image preprocessing module, an image quality assessment module, a detection and inference module, and a result determination and feedback module.

[0029] The image upload and communication module receives raw images and image identification information uploaded from the smartphone and performs integrity verification on the received image data. This verification includes checking the integrity of the data packet, the image file for corruption, and the completeness of the identification information. If data is missing or abnormal, a retake prompt is immediately sent to the smartphone, guiding the user to retake and upload the image. Integrity verification promptly eliminates invalid data, preventing subsequent processing errors due to transmission problems.

[0030] The image preprocessing module is connected to the image upload and communication module and is used to preprocess the received raw images to obtain the images to be evaluated. Preprocessing includes one or more of the following: brightness normalization, contrast normalization, white balance normalization, and resolution unification. These preprocessing operations can effectively eliminate image quality fluctuations caused by differences in shooting environment (such as lighting conditions, shooting angle, and background color) and differences in mobile phone models, providing standardized input for subsequent processing. For example, white balance normalization can correct color casts under different light sources, making the colors in the display areas closer to reality, thereby improving detection accuracy.

[0031] The image quality assessment module is connected to the image preprocessing module and is equipped with a first deep learning neural network. This first deep learning neural network is used to extract features from the image to be assessed and output the image validity and quality assessment results. In this embodiment, the image validity and quality assessment results include three states: image valid, image blurry, or no detected target. No detected target indicates that no test kit structure or color development area was detected; an image blurry indicates that the test kit was detected but the image clarity does not meet the confidence threshold. The model size of the first deep learning neural network is smaller than that of the subsequent second deep learning neural network. Using a lightweight network allows for rapid quality screening, improving system response speed, while intercepting blurry or invalid images outside the main process, avoiding resource waste and incorrect judgments.

[0032] The detection inference module is connected to the image quality assessment module and is equipped with a second deep learning neural network. This second deep learning neural network performs detection inference on the image to be assessed when the image validity and quality assessment result indicates the image is valid, and outputs the category detection result of the test kit. In this embodiment, the second deep learning neural network adopts a ConvNeXt network structure, and, considering the characteristics of the test kit's small color development area, subtle color changes, and complex background interference, optimizes the training strategy to enable the model to focus on the color development features of the test kit, thereby accurately distinguishing between negative, positive, or neutral results. The large model size ensures the precision and accuracy of the classification.

[0033] The result determination and feedback module is connected to the image quality assessment module and the detection inference module, respectively. It generates the final detection status based on the image validity and quality assessment results or category detection results, and returns the final detection status to the smartphone for display. The final detection status includes five types: image invalid, image blurred, neutral, positive, or negative. The specific correspondences are as follows: when the image quality assessment module outputs "no detection target," the final detection status is "image invalid"; when the output image is blurred, the final detection status is "image blurred"; when the output image is valid and the detection inference module outputs "negative," the final detection status is "negative"; when the output image is valid and the detection inference module outputs "positive," the final detection status is "positive"; when the output image is valid and the detection inference module outputs "neutral," the final detection status is "neutral." Additionally, the neutral result includes the judgment result when the confidence scores of the second deep learning neural network for both negative and positive judgments are below the confidence threshold. If the network outputs the category with the highest confidence score as negative or positive, its confidence score needs to be further compared with a preset confidence threshold. If the score is greater than or equal to the threshold, the final status is negative or positive; if the score is lower than the threshold, a safety downgrade mechanism is triggered, forcibly classifying the test and outputting a final status of neutral. Furthermore, the preset confidence threshold is an adjustable system parameter, and its setting and adjustment logic is as follows: the initial baseline value of this threshold is determined by plotting the Receiver Operating Characteristic (ROC) curve on the validation set and calculating the maximum value of the Youden Index. Considering the physical characteristic of the test strip's color development exhibiting a gradual transition, the empirical range of this threshold is typically set between 0.75 and 0.85, preferably 0.80. In actual deployment, this threshold can be dynamically adjusted according to the specific application scenario's indicator preferences: when the detection scenario requires high sensitivity to reduce the false negative rate, the threshold can be appropriately lowered; when the scenario requires high specificity to reduce the false positive rate, the threshold can be appropriately increased. By combining native three-class classification with an adjustable confidence threshold to create a neutral state interception design, the system can adaptively handle critical samples with different levels of color development. This completely avoids the risk of false positives or false negatives caused by the "forced classification" of deep learning models under edge features, greatly enhancing the clinical detection reliability of the system.

[0034] Example 2 This embodiment provides a test kit detection method based on smartphones and deep learning. This method can be applied to the system described in Embodiment 1, such as... Figure 1 As shown, it includes the following steps: Step S1: The smartphone captures the original image of the test kit and uploads the original image and image identification information to the server. Specifically, the user accesses the testing interface through a mini-program on their smartphone, and, guided by the interface, places the test kit to be tested within the preset shooting area. The smartphone camera then captures the original image containing the test kit and its color development area. After acquisition, the smartphone automatically packages and uploads the original image and image identification information (such as user ID, device model, acquisition time, etc.) to the server. The mini-program interface guidance helps users capture images that meet the requirements, reducing image quality issues caused by improper operation.

[0035] Step S2: The server performs an integrity check on the received raw image. If data is missing or abnormal, a retake prompt is sent to the smartphone. The server first checks whether the data packet is complete, whether the image file is corrupted, and whether the image identification information is complete. If data is missing or abnormal, a retake prompt is immediately sent to the smartphone, guiding the user to retake and upload the image. Integrity checks can promptly eliminate abnormalities such as transmission errors or file corruption, ensuring that subsequent processing is based on valid data.

[0036] Step S3: The server preprocesses the original image that has passed the integrity check to generate the image to be evaluated. Preprocessing operations include resolution normalization, white balance normalization, brightness normalization, and contrast normalization. These preprocessing operations eliminate interference from differences in the shooting environment, allowing the subsequent neural network model to focus on the features of the test kit itself.

[0037] Step S4: The first deep learning neural network is invoked to extract features from the image to be evaluated, outputting image validity and quality evaluation results, including whether the image is valid, blurry, or has no detected target. The first deep learning neural network uses the lightweight convolutional neural network architecture MobileNetV3-Small to quickly extract features and classify the image to be evaluated. The network will output three categories: no detected target means that no test kit structure or color area was detected, which is judged as "image invalid", and a retake prompt is returned; blurry image means that the test kit was detected but the image clarity is insufficient, which is judged as "image invalid", and a retake prompt is returned; valid image means that the image has passed the quality screening and can be input into the second network. Through this step, invalid images can be quickly identified and removed, avoiding meaningless calculations by the main network in the subsequent process.

[0038] When training the first deep learning neural network, the cross-entropy loss function was used as the optimization objective. The optimizer used was Adam, and the initial learning rate was set to 1×10⁻⁶. - ³, and a cosine annealing strategy is used for learning rate decay; the weight decay coefficient is 1×10. -4The training process was set to a maximum of 150 training epochs, and an early stopping mechanism was introduced. The validation set loss was used as the monitoring metric, and training was terminated early if the validation set loss did not significantly decrease within 15 consecutive epochs. The batch size during training was set to 64. The hardware environment used for training was an NVIDIA RTX 3090 GPU.

[0039] Step S5: Determine whether the image validity and quality assessment result indicates that the image is valid. If the result indicates that the image is valid, continue to step S6; otherwise, jump directly to step S7 to generate the final detection status based on the current result.

[0040] Step S6: The second deep learning neural network is invoked to perform detection inference on the image to be evaluated, and the category detection result of the test kit is output. The second deep learning neural network adopts the ConvNeXt network structure and outputs the probabilities of three categories: negative, positive, and neutral. To effectively handle critical samples and avoid the risk of misdiagnosis, the system innovatively introduces a confidence verification and security degradation mechanism: when the detection inference module outputs the category with the highest confidence as "negative" or "positive", its confidence score needs to be further compared with a preset confidence threshold. If the score is greater than or equal to the preset threshold, the final state is confirmed as "negative" or "positive"; if the score is lower than the preset threshold, that is, when the network is not confident enough in judging whether it is positive or negative, the security degradation mechanism is triggered, and it is forced to be classified and output as a "neutral" result.

[0041] When training the second deep learning neural network, the cross-entropy loss function was also used as the optimization objective. The optimizer used was Adam, and the initial learning rate was set to 1×10⁻⁶. - ³, and a cosine annealing strategy is used for learning rate decay; the weight decay coefficient is 1×10. -4 The training process was set to a maximum of 150 training epochs, and an early stopping mechanism was introduced. The validation set loss was used as the monitoring metric; training was terminated early if the validation set loss did not significantly decrease within 15 consecutive epochs. The batch size during training was set to 64. The hardware environment used for training was an NVIDIA RTX 3090 GPU. It should be noted that the first and second deep learning neural networks used the same training strategy, but they were trained independently, and their network parameters were not shared.

[0042] Step S7: Generate the final detection status based on the image validity and quality assessment results or category detection results, and return the final detection status to the smartphone for display. The final detection status includes image invalid, image blurry, neutral, positive, or negative. The specific correspondence is as follows: when step S4 outputs no detection target, the final detection status is image invalid; when step S4 outputs image blurry, the final detection status is image blurry; when step S4 outputs image valid and step S6 outputs negative, the final detection status is negative; when step S4 outputs image valid and step S6 outputs positive, the final detection status is positive; when step S4 outputs image valid and step S6 outputs neutral, the final detection status is neutral. After receiving the results, the smartphone displays the detection conclusion to the user in text and graphic form. If the result is image invalid or image blurry, the user is prompted to retake the photo; if the result is negative, positive, or neutral, the corresponding detection conclusion is displayed, and further suggestions or guidance can be provided (such as suggesting medical attention if positive, indicating normal if negative). Through clear user interaction, the usability and practicality of the detection system are improved.

[0043] Both the first and second deep learning neural networks used the same optimizer, hyperparameters, early stopping strategy, and hardware environment during training to ensure the fairness and reproducibility of the comparative experiment.

[0044] To demonstrate that the "decoupled architecture + lightweight network" scheme of this invention is superior to other schemes, this embodiment designs a comparative experiment to compare the performance of different schemes on the test kit detection task. The experiment sets up three schemes for comparison: (1) Scheme A (Decoupling and Lightweight): The decoupling scheme of the present invention is adopted. The quality assessment module uses the lightweight network MobileNetV3-Small, and the detection and inference module uses the superior ConvNeXt network.

[0045] (2) Scheme B (coupled integration): Integrate quality assessment and detection into the same neural network module, and use ConvNeXt to perform quality assessment and category detection simultaneously.

[0046] (3) Scheme C (no quality control): Only the detection module is set up, without the quality assessment module, and ConvNeXt is used directly for category detection.

[0047] The same dataset was used for both training and testing in the experiments. The dataset included clear and valid images, blurred images, images without targets, and test kit images with different color development results (negative, positive, and neutral). The experimental results are shown in Tables 1 and 2.

[0048] Table 1 Comparison of accuracy rates of the proposed solutions

[0049] Table 2. Efficiency Comparison of Schemes

[0050] The experimental results show that Scheme A, employing the decoupled lightweight approach of this invention, outperforms other schemes in overall classification accuracy, effective image interpretation accuracy, and abnormal image recognition accuracy. It also exhibits significant advantages in inference latency and resource consumption. Specifically, Scheme A, through its pre-assessment module, effectively identifies and eliminates blurry and targetless images, preventing these anomalous samples from interfering with the main classification network and thus improving overall accuracy. The use of a lightweight network ensures high processing efficiency. Scheme B, by coupling quality assessment with detection, has more parameters but slightly lower performance and higher resource consumption. Scheme C, lacking quality assessment, while offering fast processing speed, cannot handle anomalous images, resulting in the lowest overall accuracy. These data fully validate the advanced nature and practicality of the technical solution of this invention.

[0051] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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. Such 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, and should all be included within the protection scope of the present invention.

Claims

1. A test kit detection system based on smartphones and deep learning, characterized in that, include: The smartphone app is used to capture images of the test kit, upload the raw images to the server, and receive and display the test results. A server-side component, communicatively connected to the smartphone, includes: The image uploading and communication module is used to receive the original image and image identification information uploaded by the smartphone. The image preprocessing module is used to preprocess the received raw image to obtain the image to be evaluated; The image quality assessment module is equipped with a first deep learning neural network, which is used to extract features from the image to be assessed and output the image validity and quality assessment results, including whether the image is valid, blurry, or has no detected target. The detection and reasoning module is equipped with a second deep learning neural network, which is used to perform detection and reasoning on the image to be evaluated when the image validity and quality evaluation result is that the image is valid, and output the category detection result of the test kit; The result determination and feedback module is used to generate a final detection status based on the image validity and quality assessment results or the category detection results, and return the final detection status to the smartphone.

2. The test kit detection system based on smartphones and deep learning according to claim 1, characterized in that, The image preprocessing module performs one or more of the following preprocessing operations: brightness normalization, contrast normalization, white balance normalization, and resolution unification.

3. The test kit detection system based on smartphones and deep learning according to claim 1, characterized in that, The model size of the first deep learning neural network is smaller than that of the second deep learning neural network.

4. A test kit detection method based on smartphones and deep learning, characterized in that, Includes the following steps: Step S1: The smartphone captures the original image of the test kit and uploads the original image and image identification information to the server. Step S2: The server performs integrity verification on the received original image. If data is missing or abnormal, a retake prompt is sent to the smartphone. Step S3: The server preprocesses the original image that has passed the integrity check to generate the image to be evaluated; Step S4: Call the first deep learning neural network to extract features from the image to be evaluated, and output the image validity and quality evaluation results, including whether the image is valid, blurry, or has no detected target; Step S5: Determine whether the image validity and quality assessment result indicates that the image is valid; Step S6: If yes, then call the second deep learning neural network to perform detection inference on the image to be evaluated and output the category detection result of the test kit; Step S7: Generate a final detection status based on the image validity and quality assessment results or the category detection results, and return the final detection status to the smartphone for display.

5. The test kit detection method based on smartphones and deep learning according to claim 4, characterized in that, The "no target detected" output in step S4 means that no test kit structure or color development area was detected; the "blurred image" means that the test kit was detected but the clarity did not meet the confidence threshold.

6. The test kit detection method based on smartphones and deep learning according to claim 4, characterized in that, The second deep learning neural network adopts the ConvNeXt network structure.

7. The test kit detection method based on smartphones and deep learning according to claim 4, characterized in that, The final detection status generated in step S7 includes: image invalid, image blurry, neutral, positive or negative.

8. The test kit detection method based on smartphones and deep learning according to claim 7, characterized in that, The image failure is the result of the judgment when there is no detected target output in step S4; the image blurriness is the result of the judgment when the image is blurriness output in step S4.

9. The test kit detection method based on smartphones and deep learning according to claim 7, characterized in that, The neutral result includes the judgment result when the confidence of the second deep learning neural network for both negative and positive judgments is lower than the confidence threshold.

10. The test kit detection method based on smartphones and deep learning according to claim 5, characterized in that, In step S6, only the valid results of the images output in step S4 are qualitatively determined.

Images