A non-intrusive panel temperature OCR identification method and system for a refrigerated container

Abstract

The present application relates to the technical field of image recognition and intelligent monitoring, and particularly relates to a non-intrusive panel temperature OCR identification method and system for refrigerated containers. The technical scheme comprises: the method comprises collecting panel images and preprocessing, detecting temperature display areas and performing perspective distortion correction, using a deep learning sequence identification model for character recognition, obtaining temperature values and confidence, checking the identification results and uploading the background, triggering an alarm when the threshold is exceeded, and reviewing through a mobile terminal. The system is composed of a visual acquisition device, an edge computing node, a background server and a mobile terminal, supports flexible fixed assembly rapid deployment, network interruption transmission, multi-modal fusion identification and equipment recycling. The present application realizes high-precision, high-robustness, low-delay, closed-loop temperature automatic identification and alarm while keeping the container completely non-intrusive, significantly reduces the labor cost and equipment investment of cold chain monitoring, and has a broad industrial application prospect.

Description

Technical Field

[0001] This invention relates to the field of image recognition and intelligent monitoring technology, and in particular to a non-invasive panel temperature OCR recognition method and system for refrigerated containers. Background Technology

[0002] Refrigerated containers are core equipment in cold chain logistics used to transport temperature-sensitive goods. Their temperature display panels are usually located at the front or side of the container to show the real-time temperature inside. At ports, yards, and warehouses, operators need to regularly record or monitor the temperature of refrigerated containers to ensure that goods are within the prescribed temperature control range and to avoid damage to goods due to refrigeration equipment malfunctions, incorrect settings, or abnormal temperatures.

[0003] Currently, the temperature of refrigerated containers is mainly obtained through the following methods:

[0004] Staff regularly check the temperature displayed on the panel and record it manually or by entering it into a handheld terminal. This method has problems such as high labor costs, low inspection frequency, easy omissions and errors, and difficulty in operation at night or in bad weather. In addition, it cannot achieve real-time monitoring and alarms.

[0005] Installing temperature sensors inside or outside the container and uploading the data via cables or wireless modules requires modifying the container structure (such as opening holes or fixing sensor probes), which is complex, time-consuming, and may affect the container's sealing and safety. It is also not suitable for leased or short-term containers.

[0006] The system uses a regular camera to capture images of the panel and traditional optical character recognition technology to extract temperature values. However, the complex and variable environment of refrigerated containers presents the following technical challenges: temperature panels are often subject to interference from reflections, water mist, dirt, low light, or strong backlight; different brands and models of containers use various display types such as digital tubes and LCD screens, with different fonts, colors, and formats including decimal points or negative signs; and the camera installation angle or curved panel surface can easily cause perspective distortion, affecting the accuracy of character positioning and recognition.

[0007] To address this issue, we propose a non-invasive OCR method and system for identifying panel temperature in refrigerated containers. Summary of the Invention

[0008] The purpose of this invention is to address the problems existing in the background art by proposing a non-invasive panel temperature OCR recognition method and system for refrigerated containers.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a non-intrusive panel temperature OCR recognition method for refrigerated containers, comprising the following steps:

[0010] S1. Acquire images of the temperature display panel of the refrigerated container and preprocess the images;

[0011] S2. Detect the temperature display area of ​​the preprocessed image and correct the perspective distortion of the temperature display area.

[0012] S3. Perform character recognition on the calibrated temperature display area to obtain the temperature value and corresponding confidence information;

[0013] S4. Verify the temperature value and upload the verified temperature value and verification status to the backend server.

[0014] S5. When the temperature value meets the preset conditions, an alarm is triggered and displayed and verified via a mobile terminal.

[0015] Furthermore, the method also includes a non-intrusive deployment step:

[0016] The visual acquisition device is installed on the outside of the temperature display panel using a flexible fixing component, and the binding relationship between the device and the container is established by scanning the device identification information with a mobile terminal.

[0017] Furthermore, the preprocessing includes at least one image enhancement operation: noise reduction, contrast enhancement, adaptive brightness adjustment, reflection suppression, and wide dynamic range optimization.

[0018] Furthermore, the temperature display area detection is achieved using at least one of a deep learning-based target detection model and an image feature-based region localization method, and can be adapted to different types of temperature display panels, including at least one of digital tube displays and liquid crystal displays.

[0019] Furthermore, the perspective distortion correction includes:

[0020] The corner positions of the temperature display area are determined based on edge and key point information in the image, and a front view image is obtained through geometric transformation.

[0021] Furthermore, the character recognition is achieved using a deep learning-based sequence recognition model, which is compatible with at least one temperature display panel in a refrigerated container, such as a digital tube or an LCD panel, and can recognize temperature information containing numbers, decimal points, and negative signs.

[0022] Furthermore, the method also includes a multimodal fusion recognition step: performing at least one of semantic verification and context verification on the character recognition result, and outputting a temperature value with confidence information.

[0023] Furthermore, the result verification includes at least one of format verification, numerical range verification, and continuous frame timing consistency verification; when at least one of the following occurs: verification failure or confidence level is lower than a preset threshold, a local re-identification operation is performed, and after multiple identification failures, at least one of the following is triggered: anomaly marking, manual review, or device status anomaly alarm.

[0024] A non-invasive panel temperature OCR recognition system for refrigerated containers includes:

[0025] Visual acquisition equipment used to acquire images of temperature display panels;

[0026] Edge computing nodes are used to perform image preprocessing, region detection, perspective distortion correction, character recognition, multimodal fusion recognition and result verification, while also enabling local caching of temperature data and resumption of data transmission after network outages.

[0027] The backend server is used to receive structured data corresponding to temperature values ​​and perform storage, analysis, and alarm processing.

[0028] The mobile terminal is used to enable device binding, temperature display, anomaly verification, real-time viewing of the temperature display panel video, and manual correction of temperature values.

[0029] The modules are connected to each other through at least one of wired communication and wireless communication methods and corresponding communication protocols.

[0030] Furthermore, the visual acquisition device includes a rapid deployment structure, which includes:

[0031] A flexible fixing component for fixing the visual acquisition device to the outside of the container temperature panel, an integrated connection component for power supply and data transmission, and a mobile terminal binding module for establishing the correspondence between the device and the container.

[0032] The system supports rapid installation and disassembly without tools and can be reused among multiple containers.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] The visual acquisition device is directly installed on the outside of the temperature display panel using flexible fixing components, without any tools, holes in the container or modifications. The device can be reused infinitely in different containers and can be untied with one click when leaving the container, which greatly reduces the monitoring cost per container.

[0035] Employing a deep learning-based target detection model and sequence recognition model, trained on a large-scale dataset including digital tubes, LCD screens, different brands / fonts, and those with decimal points and negative signs, it can adaptively detect and recognize the temperature information of various refrigerated container panels without the need to manually configure parameters for different panel types.

[0036] Corner detection and perspective transformation are used to perform geometric correction on the temperature display area to eliminate deformation caused by camera mounting angle or panel curvature. At the same time, multi-level preprocessing such as noise reduction, adaptive histogram equalization, adaptive brightness adjustment, wide dynamic range optimization and deep learning-based reflection repair are used to effectively suppress environmental interference such as reflection, water stains and low light, and significantly improve the success rate of subsequent character recognition.

[0037] The edge nodes first perform lightweight OCR recognition. For images with low confidence or failed verification, they automatically trigger secondary recognition using a multimodal large model in the background. The results from both methods are then fused to output a high-confidence temperature value. Combined with format verification, numerical range verification, and continuous frame temporal consistency verification, this effectively filters out misidentifications and jump anomalies. When multiple recognition failures occur, manual review is automatically pushed, forming a closed loop of "machine recognition + large model assistance + manual confirmation" to meet industrial-grade reliability requirements.

[0038] All image preprocessing, region detection, correction, recognition, and verification tasks are completed locally on the edge computing node, with a single frame processing time of less than 500ms. It does not rely on real-time cloud inference, avoiding the impact of network latency or jitter. Simultaneously, it has a built-in local data cache that automatically stores recognition results when the network is interrupted and resumes transmission after the network connection is restored, ensuring no data loss.

[0039] The backend server has preset temperature threshold rules. When the threshold is exceeded, it automatically generates a graded alarm and pushes the service to the mobile terminal of the management personnel. The mobile terminal supports real-time video streaming, AI recognition of temperature overlay display, abnormal image viewing and manual temperature correction, so that on-site personnel can quickly identify abnormal situations and take measures to reduce the risk of cargo damage.

[0040] The system supports PoE wired connection and 4G / 5G / WiFi wireless communication, and can be flexibly deployed according to the network conditions of the yard; the visual acquisition equipment has an IP67 protection rating and an anti-fog protective cover, and is suitable for harsh outdoor environments such as rain, snow, high humidity, and dust. Attached Figure Description

[0041] Figure 1 This is a system block diagram of the present invention;

[0042] Figure 2 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0044] Example 1

[0045] like Figure 1 As shown, this embodiment provides a non-intrusive panel temperature OCR recognition system for refrigerated containers, which consists of the following four core components:

[0046] Visual acquisition equipment: The visual acquisition equipment is a 2-megapixel starlight-level PoE industrial camera with an IP67 protection rating, and a built-in infrared supplementary lighting module and photosensor. A custom-designed waterproof and anti-fog protective cover (ABS+PC material) is installed on the outside of the camera. The front window of the cover uses optically coated glass to effectively reduce condensation and reflection. The camera is quickly and easily secured to the outside of the container's temperature panel using a spring-loaded cable, and supports omnidirectional angle adjustment to ensure the panel is centered in the image.

[0047] Edge computing nodes: Edge computing nodes adopt ARM architecture embedded motherboards (such as RK3588), integrating an NPU (6TOPS computing power), 8GB of memory, and 64GB of storage. These nodes can be integrated inside the visual acquisition device housing or deployed independently in an explosion-proof low-voltage box and connected to the camera via network cable. The nodes deploy the following software modules: image preprocessing module (denoising, enhancement, reflection suppression); temperature display area detection module (YOLOv8); perspective distortion correction module (corner detection + perspective transformation); character recognition module (CRNN+CTC); multimodal fusion recognition module (lightweight verification based on rule and confidence fusion, automatically triggering secondary verification of the large background model when the confidence level is below a threshold); result verification and local re-recognition module; data caching and network interruption resumption module (SQLite).

[0048] Edge nodes are directly connected to the camera via PoE cable and simultaneously connected to the backend server via wired or 4G / 5G wireless means. The nodes perform all AI inference locally (except for secondary verification of the large model) and only upload structured data (temperature value, confidence level, timestamp, device ID, anomaly flag). The local single frame processing time is <500ms.

[0049] Backend server: The backend server is deployed in the cloud or on-premises data center and includes the following functional modules:

[0050] Device Binding Management Module: Maintains the mapping relationship between device IDs and container numbers, and supports binding, unbinding, and status query via QR code scanning;

[0051] Temperature data storage and visualization module: Temperature data is stored using a time-series database (such as InfluxDB), and real-time and historical curves are displayed on a 3D monitoring screen;

[0052] Large Model Secondary Verification Module: Deploys the QWEN2.5 multimodal large model, receives low-confidence or verification-failed images uploaded by edge nodes, performs deep recognition, and returns the results;

[0053] Alarm engine and message push module: Preset temperature threshold rules (such as below -25℃ or above -10℃). When the threshold is exceeded, a graded alarm is generated and sent to the mobile terminal via WebSocket or JPush.

[0054] Mobile terminals:

[0055] The mobile terminal is a smartphone or tablet with a dedicated app installed, and has the following functions:

[0056] QR code binding module: Scan the device's serial number and container number to complete the binding relationship creation;

[0057] Real-time video streaming and AI temperature overlay module: pulls the camera RTSP stream, decodes it and displays the panel screen in real time, and overlays the AI-recognized temperature values ​​returned by the edge nodes;

[0058] Manual verification and temperature correction module: For abnormal markers or low-confidence identifications, the APP can display real-time video or cached images on site, which can be manually read and confirmed by operators; manual correction and re-reporting are supported.

[0059] Equipment recycling and unbinding module: Unbind the equipment with one click when it leaves the factory, and restore the equipment to an idle state.

[0060] The visual acquisition device communicates with the edge computing node via PoE (power and data transmission simultaneously); the edge computing node connects to the backend server via at least one communication protocol, such as wired Ethernet, 4G / 5G, or WiFi; and the mobile terminal connects to the backend server via 4G / 5G or WiFi. Data exchange between modules is performed using JSON format and MQTT / HTTPS protocols.

[0061] Through the coordinated operation of the above four parts, the system achieves a fully automated closed loop from image acquisition, edge recognition, background storage and alarm to mobile terminal review and retrieval.

[0062] Example 2

[0063] like Figure 2As shown, this embodiment provides a non-intrusive panel temperature OCR recognition method for refrigerated containers. This method is executed collaboratively by edge computing nodes and a backend server, and specifically includes the following steps:

[0064] S1. Acquire and preprocess images.

[0065] First, the original image of the temperature display panel of the refrigerated container is acquired in real time using a visual acquisition device (such as a 2-megapixel starlight-level POE industrial camera with IP67 protection rating). The camera is installed about 20-50cm outside the temperature panel, with the viewing angle directly facing the display area.

[0066] The acquired raw images may be affected by uneven lighting, reflections, dirt, water stains, etc. Therefore, at least one of the following preprocessing operations should be performed:

[0067] Denoising: Gaussian filtering or bilateral filtering algorithms are used to remove sensor noise and environmental noise from the image to avoid interfering with subsequent region detection.

[0068] Contrast Enhancement: Apply Adaptive Histogram Equalization (CLAHE) to improve the contrast between characters and background, especially suitable for low-light or backlit scenes.

[0069] Adaptive brightness adjustment: Based on global image brightness statistics, it dynamically adjusts exposure compensation or gamma correction to maintain appropriate brightness in the temperature display area.

[0070] Reflection suppression: When a reflective area of ​​the glass panel is detected, multiple frames of images with different exposure parameters are acquired to synthesize a wide dynamic range (WDR) image. Then, a mask of the reflective area is output using a trained U-Net reflective detection network. Finally, the fast traversal method is used to perform image inpainting within the mask to restore the characters that were covered by the reflection.

[0071] Wide dynamic range optimization: For scenes with strong direct light, the dynamic range is expanded by multi-frame fusion, while preserving highlights and shadow details.

[0072] In a preferred embodiment, the preprocessing sequentially performs noise reduction, CLAHE contrast enhancement, adaptive brightness adjustment, and reflection suppression to cope with the complex lighting environment of outdoor storage yards.

[0073] S2. Temperature display area detection and perspective distortion correction

[0074] The preprocessed image is input to the temperature display area detection module. In this embodiment, a deep learning-based object detection model (such as YOLOv8) is used to locate the temperature display area. This model is pre-trained on a dataset of refrigerated container panels containing various brands (such as Maersk, CIMC, Taicang CIMC, etc.) and various display types (digital tube, LCD screen, with decimal point or minus sign). It can output the bounding box and confidence score of the temperature display area. Alternatively, a region localization method based on image features can be used, such as using the texture features of character regions (MSER or SWT) combined with morphological operations to extract candidate regions.

[0075] After detecting the temperature display area, perspective distortion correction is performed. Because the camera mounting angle or panel curvature may cause geometric distortion of the image, affecting character recognition accuracy, the correction steps are as follows:

[0076] Within the detected area, edges are extracted using the Canny edge detection operator, and then four straight lines are detected using the Hough transform, corresponding to the top, bottom, left, and right edges of the panel, respectively.

[0077] Calculate the perspective transformation matrix (homography matrix) based on the correspondence between the four corner points and the target front view rectangle (e.g., set to a width of 300 pixels and a height of 100 pixels).

[0078] A perspective transformation is applied to the original image, and bilinear interpolation is performed for resampling to obtain a corrected front-view rectangular image in which the character rows and columns remain horizontal and vertical.

[0079] When line detection fails due to reasons such as panel edge occlusion or extremely poor lighting, the system automatically switches to an alternative solution: directly using the four vertices of the bounding box output by YOLOv8 as corner points to calculate the perspective transformation matrix, or using a contour-based corner detection algorithm (such as Douglas-Peucker polygon approximation) to extract valid corner points. If the alternative solution still fails, the frame is marked as "region detection failed," triggering local re-identification (retrying after adjusting image enhancement parameters) or reporting an anomaly.

[0080] S3. Character recognition obtains temperature values ​​and corresponding confidence information.

[0081] The corrected temperature display area image is input into the character recognition module. In this embodiment, a deep learning-based sequence recognition model—CRNN+CTC (convolutional recurrent neural network + connection-time classification)—is used to achieve end-to-end temperature value recognition. CRNN automatically extracts the image feature sequence, and CTC decodes and outputs the character sequence, which can recognize numbers (0-9), decimal point (.), and minus sign (-), adapting to different fonts of digital tubes and LCD panels.

[0082] To further improve recognition robustness, this embodiment also includes a multimodal fusion recognition step:

[0083] The edge node first performs CRNN recognition to obtain preliminary temperature values ​​and confidence levels. If the confidence level is ≥0.85, it directly proceeds to step S4. If the verification passes, the result is output; if the verification fails, local re-recognition is performed. If the confidence level is <0.85, the image is uploaded to the backend large model for secondary recognition. After the large model returns the results, the result is then verified.

[0084] The QWEN2.5 multimodal large model deployed on the backend server performs secondary recognition on the same image region and performs semantic verification (e.g., checking whether "-" appears reasonably and whether the decimal point position is legal) and context verification (e.g. comparing whether the temperature values ​​of the previous and next frames are continuous).

[0085] Finally, the results of CRNN and the large backend model are merged to output a temperature value with a comprehensive confidence level: if the two identification results are consistent and both have high confidence levels, then the result with higher confidence level is adopted; if there is a discrepancy, then the result with higher confidence level is selected; if the large backend model cannot identify the value, then manual review is triggered.

[0086] S4. Result Verification and Upload

[0087] The identified temperature value needs to undergo multiple verifications to be considered valid. This embodiment performs at least one of the following verifications:

[0088] Format validation: Check if the temperature string conforms to the regular expression "[-]?[0-9]+(.[0-9]+)?" and exclude illegal characters.

[0089] Numerical range verification: Determines whether the temperature is within a reasonable range, such as -50℃ to +70℃. If it exceeds the range, it is marked as abnormal.

[0090] Continuous frame timing consistency check: Compare the temperature identified in the current frame with the historical temperature of the previous 3 frames, calculate the rate of change. If the current frame is one of the previous 3 frames after the system starts, skip the timing check or only compare with the existing frames. If the rate of change exceeds the preset mutation threshold (e.g., 5℃ / second), it is determined to be an identification error or panel jump, triggering re-identification.

[0091] If all verifications pass, the temperature value, confidence level, timestamp, device ID, and verification status ("valid") are encapsulated in JSON format and uploaded to the backend server via the MQTT protocol. If any verification fails or the confidence level is lower than a preset threshold (e.g., 0.85), a local re-identification operation is performed: adjusting image enhancement parameters (e.g., increasing contrast intensity, changing the reflection repair radius) or repositioning the temperature display area, repeating steps S2 to S4 up to 3 times. If it still fails, the frame is marked as "identification abnormal," the abnormal frame image is cached locally, and pushed to the mobile terminal for manual review.

[0092] In addition, edge computing nodes perform periodic self-checks: if image acquisition fails 10 times in a row, NPU inference times out, or the camera goes offline, an "abnormal device status alarm" is generated and uploaded to the backend server, which then forwards it to the mobile terminal.

[0093] S5, Alarms and Mobile Verification

[0094] The backend server receives temperature data in real time, stores it in a time-series database, and displays current and historical temperature curves on a 3D monitoring dashboard. The system presets temperature thresholds (e.g., triggering alarms when below -25℃ or above -10℃). When the temperature value meets preset conditions (e.g., exceeding the threshold for more than 10 seconds), the alarm engine generates a tiered alarm (Level 1: Severe, Level 2: Warning) and pushes it to the mobile terminal APP of relevant administrators via WebSocket.

[0095] The mobile app has the following functions:

[0096] Device binding interface: Scan the code to complete the binding of the device and the container.

[0097] Real-time video streaming: Pull the camera's RTSP stream and overlay it with AI-recognized real-time temperature values.

[0098] Temperature Confirmation and Correction: Users can confirm and report the temperature identified by AI with one click; if an error is found, users can manually enter the correct temperature and submit.

[0099] Anomaly Verification: When an anomaly marker or alarm is received, the APP displays the corresponding anomaly frame image or real-time video, and the user can manually determine whether the temperature is indeed abnormal.

[0100] Example 3: Non-intrusive Deployment and Quick Binding

[0101] Workers retrieved the AI ​​vision acquisition device from the storage cabinet. (This device integrates a camera, supplementary lighting, and edge computing nodes, which can be built into the device housing or connected as a standalone unit via quick-connect connectors.) The back of the device is secured with a spring-loaded tension rope (covered with a silicone protective layer), a rotatable universal camera bracket, and a magnetic anti-drop auxiliary base. Operators, without any tools, simply stretch the spring-loaded tension rope and hook it onto the outer edge of the container's temperature panel or a ventilation hole. Releasing it automatically locks the rope in place, ensuring the camera lens is directly facing the temperature display area. The universal bracket allows for fine-tuning of the tilt and rotation angles to ensure the panel is centered in the image.

[0102] The device is connected to the PoE port of the field industrial switch via a Cat6e shielded network cable (PoE power supply), which simultaneously enables power supply and gigabit Ethernet data transmission. If the edge computing node is deployed independently, the camera is connected to the edge node via PoE, and the edge node is then connected to the switch via another network port.

[0103] The operator opens the mobile app, clicks "Add Device," scans the serial number (SN) QR code on the device's casing, and then scans the container number barcode. The app sends the device ID and container number to the backend server. The backend verifies whether the device is in an "idle" state. If there is no conflict, a binding relationship is established, and the camera video stream address and recognition parameters (such as ROI coordinates and enhancement mode) are sent to the edge node. After binding is complete, the app displays "Bound" and the operator can directly view the live video.

[0104] When the container leaves the site, the operator clicks "Unbind and Recycle" on the app. The system then disconnects the device from the container, and the device status returns to "idle." The staff releases the spring tension rope, removes the device, and places it in the recycling bin, ready to be bound again for use when the next container arrives. The entire process requires no disassembly of any screws or damage to the container structure.

[0105] Example 4: Adaptive Recognition of Multi-Brand Panels

[0106] Because the temperature panels of refrigerated containers vary between different brands, for example:

[0107] Maersk cold boxes: red digital tubes, black background, and large character spacing;

[0108] CIMC Cold Box: Blue LCD screen with one decimal place and a minus sign displayed on the left;

[0109] Other models: Other brands of cold boxes, the characters have broken strokes.

[0110] This invention employs the following strategy to achieve adaptivity:

[0111] Training data diversity: Collected images of cold box panels from more than 50 different brands / models, labeled the temperature numerical regions, and constructed a dataset containing 100,000 images.

[0112] Model structure: The YOLOv8 detection head outputs categories including "digital tube area" and "LCD area" to assist the classifier in determining the panel type.

[0113] Character recognition preprocessing: Automatically select different binarization thresholds (fixed high threshold for digital tubes, adaptive threshold for LCD screens) and morphological operations (digital tubes require expansion connection breakpoints, LCD screens do not).

[0114] Post-recognition processing: For the minus sign on the LCD screen, it may be displayed as "-" or "—". It is uniformly mapped to the standard minus sign by the aspect ratio of the character.

[0115] Example 5: Specific Implementation of the Rapid Deployment Structure

[0116] The rapid deployment architecture includes the following components:

[0117] Flexible fixing components: Spring-loaded tension rope, silicone-coated, adjustable length from 20-60cm, with anti-slip hooks at both ends. The tension rope passes through a slot at the bottom of the universal bracket. The bracket base has a soft magnetic plate for adhering to the surface of a metal box for auxiliary fixing.

[0118] Integrated connection components: A waterproof RJ45 interface extends from the rear of the camera, connecting to a PoE switch via a Cat6e shielded network cable. The network cable is pre-lengthed to 3 meters and comes with a spiral winder to prevent it from being too loose or too tight.

[0119] Mobile binding module: The app has a built-in barcode scanning library that supports QR code, barcode, and RFID reading. Binding relationships are stored in a background Redis cache and synchronized periodically by edge nodes.

[0120] This structure allows a single device to be used from pickup to installation and binding in less than 1 minute on average, and it can be reused an unlimited number of times between different containers.

[0121] The above specific embodiments are merely several further embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

[0122] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

Claims

1. A non-invasive panel temperature OCR recognition method for refrigerated containers, characterized in that, Includes the following steps: S1. Acquire images of the temperature display panel of the refrigerated container and preprocess the images; S2. Detect the temperature display area of ​​the preprocessed image and correct the perspective distortion of the temperature display area. S3. Perform character recognition on the calibrated temperature display area to obtain the temperature value and corresponding confidence information; S4. Verify the temperature value and upload the verified temperature value and verification status to the backend server. S5. When the temperature value meets the preset conditions, an alarm is triggered and displayed and verified via a mobile terminal.

2. The non-invasive panel temperature OCR recognition method for refrigerated containers according to claim 1, characterized in that: The method also includes a non-intrusive deployment step: The visual acquisition device is installed on the outside of the temperature display panel using a flexible fixing component, and the binding relationship between the device and the container is established by scanning the device identification information with a mobile terminal.

3. The non-invasive panel temperature OCR recognition method for refrigerated containers according to claim 1, characterized in that: The preprocessing includes at least one image enhancement operation: noise reduction, contrast enhancement, adaptive brightness adjustment, reflection suppression, and wide dynamic range optimization.

4. The non-intrusive panel temperature OCR recognition method for refrigerated containers according to claim 1, characterized in that: The temperature display area detection is achieved using at least one of a deep learning-based target detection model and an image feature-based region localization method, and can be adapted to different types of temperature display panels, including at least one of digital tube displays and liquid crystal displays.

5. The non-invasive panel temperature OCR recognition method for refrigerated containers according to claim 1, characterized in that: The perspective distortion correction includes: The corner positions of the temperature display area are determined based on edge and key point information in the image, and a front view image is obtained through geometric transformation.

6. The non-invasive panel temperature OCR recognition method for refrigerated containers according to claim 1, characterized in that: The character recognition is achieved using a deep learning-based sequence recognition model, which is compatible with at least one temperature display panel in refrigerated containers, such as digital tubes or LCD panels, and can recognize temperature information containing numbers, decimal points, and negative signs.

7. The non-invasive panel temperature OCR recognition method for refrigerated containers according to claim 1, characterized in that: The method further includes a multimodal fusion recognition step: performing at least one of semantic verification and context verification on the character recognition result, and outputting a temperature value with confidence information.

8. The non-invasive panel temperature OCR recognition method for refrigerated containers according to claim 1, characterized in that: The result verification includes at least one of format verification, numerical range verification, and continuous frame timing consistency verification. When at least one of the following occurs: verification failure or confidence level is lower than a preset threshold, a local re-identification operation is performed, and after multiple identification failures, at least one of the following is triggered: an anomaly marker, manual review, or device status anomaly alarm.

9. A non-intrusive panel temperature OCR recognition system for refrigerated containers, characterized in that, include: Visual acquisition equipment used to acquire images of temperature display panels; Edge computing nodes are used to perform image preprocessing, region detection, perspective distortion correction, character recognition, multimodal fusion recognition and result verification, while also enabling local caching of temperature data and resumption of data transmission after network outages. The backend server is used to receive structured data corresponding to temperature values ​​and perform storage, analysis, and alarm processing. The mobile terminal is used to enable device binding, temperature display, anomaly verification, real-time viewing of the temperature display panel video, and manual correction of temperature values. The modules are connected to each other through at least one of wired communication and wireless communication methods and corresponding communication protocols.

10. A non-intrusive panel temperature OCR recognition system for refrigerated containers according to claim 9, characterized in that: The visual acquisition device includes a rapid deployment structure, which includes: A flexible fixing component for fixing the visual acquisition device to the outside of the container temperature panel, an integrated connection component for power supply and data transmission, and a mobile terminal binding module for establishing the correspondence between the device and the container. The system supports rapid installation and disassembly without tools and can be reused among multiple containers.

Images