Intelligent pet door based on image recognition and control method thereof

Abstract

The invention provides an intelligent pet door based on image recognition and a control method thereof, and the intelligent pet door comprises an image collection device which is used for obtaining an image of a pet door region; the artificial intelligence processing unit is used for analyzing the image by adopting a deep learning algorithm and / or a machine learning algorithm, and executing the following functions: identifying the individual identity of the pet by analyzing the appearance characteristics of the pet; detecting and / or analyzing the mouth area of the pet, wherein the analysis of the mouth area of the pet comprises any one or more of the following steps: detecting whether foreign matters or articles exist in the mouth area, identifying or classifying the foreign matters or articles in the mouth area, and analyzing the form or state change of the mouth area; and the door control device controls the passing state of the pet door according to the analysis result of the artificial intelligence processing unit. The key problem that an existing pet door cannot intelligently recognize objects carried in the mouth of a pet and control the door lock according to the objects is solved, meanwhile, the functions of individual recognition and health state monitoring are integrated, and intelligent and precise pet access management is achieved.

Description

Technical Field

[0001] This invention relates to the field of smart home device technology, and in particular to a smart pet door based on image recognition and its control method. Background Technology

[0002] Pet doors are a common feature in modern homes, especially for families with cats, dogs, and other pets. They provide pets with the convenience of freely entering and exiting specific areas while minimizing disruption to the owner's life. Traditional pet doors are mainly divided into two categories: mechanical pet doors and electronic pet doors.

[0003] Mechanical pet gates typically use one-way or two-way swing doors, relying on the pet's own pushing force to pass through. While simple in structure, they lack selective control, allowing any animal of a certain size to enter. Electronic pet gates, on the other hand, achieve basic access control through technological means, most commonly based on radio frequency identification (RFID) technology. In this system, the pet wears a collar containing a specific electronic tag. When the pet approaches, a reader on the gate recognizes the valid tag and unlocks the door. Electronic pet gates improve security but still have significant limitations: First, they cannot identify the pet itself, only the collar. If the collar is lost, damaged, or worn by another animal, the system will malfunction or misjudge. Second, they cannot perceive the pet's behavioral state, especially whether the pet is carrying anything in its mouth.

[0004] For cat owners, a common concern is cats' natural hunting instincts. Cats frequently bring home prey such as birds and rodents, which can cause household hygiene problems, health risks for both pets and owners, and negatively impact local wildlife conservation. Currently, there are no commercially available pet door products that effectively address this issue of preventing cats from bringing prey indoors. Existing technology completely lacks the capability for real-time, intelligent visual analysis of the pet's mouth area.

[0005] In recent years, computer vision and deep learning technologies have made groundbreaking progress. Convolutional neural networks and other technologies have demonstrated near-human performance in tasks such as object detection, image classification, face recognition, and pose estimation. These technological advancements have laid a solid foundation for developing more intelligent perception systems. Simultaneously, the continuous improvement in the performance of edge computing devices has made it possible to run complex deep learning models on power- and cost-constrained embedded devices. This provides a feasible technical path for integrating advanced AI vision technologies into home IoT devices.

[0006] Therefore, there is an urgent need for an intelligent pet gate and its control method that can comprehensively utilize modern image recognition and artificial intelligence technologies, especially the ability to accurately detect and classify items carried in a pet's mouth, thereby achieving intelligent and refined access management, to fill the gap in existing technologies. Summary of the Invention

[0007] To address at least one of the technical problems mentioned above, this invention provides an intelligent pet door based on image recognition and its control method. This solves the key problem that existing pet doors cannot intelligently identify objects carried in a pet's mouth and control the door lock accordingly. It also integrates individual identification and health status monitoring functions, achieving intelligent and precise pet access management.

[0008] The present invention solves the technical problem by adopting the following technical solution:

[0009] A smart pet gate based on image recognition includes:

[0010] Image acquisition device for acquiring images of the pet gate area;

[0011] The artificial intelligence processing unit uses deep learning algorithms and / or machine learning algorithms to analyze images and performs the following functions:

[0012] (a) Pet identification: Identifying individual pets by analyzing their physical characteristics;

[0013] (b) Mouth area analysis: Detection and / or analysis of the pet's mouth area, which includes any one or more of the following: detecting the presence of foreign objects or items in the mouth area, identifying or classifying foreign objects or items in the mouth area, and analyzing changes in the morphology or state of the mouth area;

[0014] The door control device controls the passage status of the pet door based on the analysis results of the artificial intelligence processing unit.

[0015] Furthermore, the identification or classification function in the mouth area analysis will distinguish the detected foreign objects or items into at least two categories: those that should be refused passage and those that are allowed passage. The artificial intelligence processing unit will send corresponding control commands to the door control device based on the classification results.

[0016] Furthermore, the deep learning algorithm is any type of neural network algorithm, including convolutional neural networks, recurrent neural networks, or transformer networks; the machine learning algorithm is any type of pattern recognition algorithm or classification algorithm.

[0017] Furthermore, the artificial intelligence processing unit also performs pet status analysis functions, which include any one or more of the following: assessing pain status by analyzing pet facial features, assessing health status by analyzing appearance features, and assessing emotional or mental state by analyzing behavioral features.

[0018] Furthermore, the mouth region analysis employs any type of image analysis technique, including any one or more of object detection, image segmentation, feature extraction, morphological analysis, anomaly detection, or semantic analysis.

[0019] Furthermore, pet personal identification is achieved by analyzing the pet's physical characteristics, which include any one or more of facial features, body shape features, coat color and pattern, body markings, or other identifiable features.

[0020] A method for controlling an intelligent pet gate based on image recognition includes the following steps:

[0021] Images of the pet gate area are acquired using an image acquisition device;

[0022] The image is analyzed by an artificial intelligence processing unit to perform pet identification and mouth area analysis;

[0023] Based on the analysis results of the mouth area, the passage status of the pet door is controlled by the door control device.

[0024] Furthermore, the specific steps of oral region analysis include:

[0025] The acquired images are preprocessed;

[0026] A deep learning object detection network is used to infer the preprocessed image and output a set of detection results including candidate bounding boxes, confidence scores, and categories.

[0027] Based on the detected pet head area, the mouth area of ​​interest is determined according to a preset ratio;

[0028] Filter the detection boxes located within the region of interest in the mouth and classified as foreign objects or items;

[0029] Based on the screening results, determine whether there are any items in the mouth. If so, further classify the items.

[0030] Furthermore, the step of determining whether there is an item in the mouth based on the screening results includes: performing a time-series voting on the judgment results of multiple consecutive frames of images, and only when the proportion of judgments that there is an item or that should be refused passage reaches a preset threshold, is the final judgment to refuse passage made.

[0031] Furthermore, the control method also includes a step of analyzing the pet's state through an artificial intelligence processing unit, including a pain state assessment; the step of the door control device controlling the pet's passage through the door also incorporates the results of the pet's state analysis.

[0032] The beneficial effects of this invention are:

[0033] (1) For the first time, real-time analysis of the pet's mouth area based on image recognition technology has been realized in pet door products. It can effectively detect whether the pet is carrying items, fundamentally solving the problem of cats bringing prey into the house, improving the level of household hygiene and helping to protect the ecological environment.

[0034] (2) It can not only detect the presence or absence of items, but also further identify the type of items and intelligently classify them into allowed and denied categories. This makes the control logic more refined and user-friendly.

[0035] (3) Pet identification is performed through visual features, eliminating the need for pets to wear any physical tags. This avoids problems such as lost, damaged, uncomfortable collars or being used by other animals, making the identification method more natural and reliable.

[0036] (4) It seamlessly integrates multiple functions such as access control, individual identification, and health monitoring into a single system. This not only provides security value but also extends the dimension of pet care, making the pet gate an intelligent pet health management node.

[0037] (5) It deeply integrates the most cutting-edge deep learning technology and proposes specific training and inference optimization strategies for real-world challenges, ensuring the accuracy and reliability of the system in practical applications.

[0038] (6) The system supports multiple deployment modes, including local, cloud, or collaborative computing, and can adapt to the different needs of different users for cost, privacy, response speed and network conditions. Attached Figure Description

[0039] Figure 1 This is a system architecture diagram of the present invention.

[0040] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0042] refer to Figure 1 This invention provides an image recognition-based smart pet gate, comprising:

[0043] An image acquisition device, which is a camera installed on the pet door, is used to acquire images of the pet door area.

[0044] The artificial intelligence processing unit uses deep learning algorithms and / or machine learning algorithms to analyze images and performs the following functions:

[0045] (a) Pet identification: Identifying individual pets by analyzing their physical characteristics;

[0046] (b) Mouth area analysis: Detection and / or analysis of the pet's mouth area, which includes any one or more of the following: detecting the presence of foreign objects or items in the mouth area, identifying or classifying foreign objects or items in the mouth area, and analyzing changes in the morphology or state of the mouth area;

[0047] The door control device controls the passage status of the pet door based on the analysis results of the artificial intelligence processing unit.

[0048] Further optimize the technical solution and analyze the images, including triggering continuous shooting when the image acquisition device approaches the door area to obtain a frame sequence. Where t is the current time or the time point corresponding to any frame (t represents the acquisition time of a certain frame in the frame sequence), and T is the preset time length or time interval, used to limit the range of the entire frame sequence.

[0049] Perform the following for each frame: size scaling to the network input resolution, pixel normalization (e.g., mapping to...). (Or perform mean and variance standardization), with optional noise reduction / sharpening.

[0050] Further optimize the technical solution and use a three-dimensional method to detect the presence of foreign objects or items in the oral region:

[0051] In each frame image Run an object detection network (e.g., YOLO11) and output a set of candidate boxes:

[0052]

[0053] in: For the position and size of the bounding box; Confidence level; Category (e.g., cat, mouth-held object / foreign object, etc.).

[0054] Using a two-stage approach—first detecting the cat's face / head, then detecting the object held in its mouth within the head's ROI—can significantly reduce false detections. The mouth ROI can be cropped proportionally to the head frame. (This is equivalent to taking the cat's face at a certain time frame t and getting the image of the cat's mouth at a certain time frame t).

[0055] When the object / foreign object category box is detected satisfy ( (The threshold for minimum duration) and the intersection-union ratio with the oral ROI ( When the crossover ratio (CROI) between the detection bounding box of the object / foreign object category and the ROI of the mouth is reached, it is determined that there is an object in the mouth.

[0056] Further optimize the technical solution. The identification or classification function in the mouth area analysis will distinguish the detected foreign objects or items into at least two categories: those that should be refused passage and those that are allowed passage. The artificial intelligence processing unit will send corresponding control commands to the door control device based on the classification results.

[0057] For example, let the set of item subcategories be... Mapping function:

[0058]

[0059] Typically, `prey / unknown` is mapped to `Reject`, and `toy / leaf`, etc., are mapped to `Allow` (you can configure this according to your product strategy). `C` is the set of item subcategories, and `f` is the mapping function that maps item subcategories to target categories that the system is interested in.

[0060] If you are currently only classifying whether someone is carrying something, it fully satisfies the requirement to at least distinguish between categories that should be refused passage and categories that are allowed passage:

[0061] Carrying something = Reject;

[0062] Uncarried item = Allow.

[0063] The degree of mouth opening can be estimated by mouth key points / segmentation (such as the distance between key points of the upper and lower lips), or the change of mouth opening / closing can be judged by continuous frame difference, which can be used to improve the robustness of object picking judgment (for example, if object picking is detected in 3 consecutive frames, it is then judged as Reject).

[0064] To further optimize the technical solution, the deep learning algorithm is any type of neural network algorithm, including convolutional neural networks, recurrent neural networks, or transformer networks; the machine learning algorithm is any type of pattern recognition algorithm or classification algorithm.

[0065] To further optimize the technical solution, the mouth region analysis adopts any type of image analysis technology, including any one or more of the following: target detection, image segmentation, feature extraction, morphological analysis, anomaly detection, or semantic analysis.

[0066] To further optimize the technical solution, the artificial intelligence processing unit also performs pet status analysis functions, which include any one or more of the following: assessing pain status by analyzing pet facial features, assessing health status by analyzing appearance features, and assessing emotional or mental state by analyzing behavioral features.

[0067] Pain status assessment based on FGS implementation: Using the FGS standard, veterinarians score cat facial images to obtain a supervised dataset, and then train a model to output the degree of pain / whether there is pain.

[0068] Let FGS be labeled as discrete level. (Or, depending on your annotation method), the model outputs a logits vector u, which is then processed by softmax to obtain the probability:

[0069]

[0070] Among them, u j The original output vector of the model represents the j-th component in the logits vector, p j This represents the probability value of the j-th FGS discrete level. k is the total number of FGS discrete levels, and j represents the category index variable.

[0071] Training using cross-entropy loss:

[0072]

[0073] in, For one-hot tags.

[0074] During inference, the level can be mapped to a binary pain / no pain rating:

[0075]

[0076] in, The pain threshold can be determined using the validation set.

[0077] To further optimize the technical solution, the mouth region analysis adopts any type of image analysis technology, including any one or more of the following: target detection, image segmentation, feature extraction, morphological analysis, anomaly detection, or semantic analysis.

[0078] Further optimization of the technical solution allows for pet personal identification by analyzing the pet's appearance features, which include any one or more of facial features, body shape features, coat color and pattern, body markings, or other identifiable features.

[0079] Pet identity can be identified using feature vector embedding and similarity matching.

[0080] The detected cat face / head ROI is input into the feature extraction network. This yields the embedding vector:

[0081]

[0082] Where ROI represents the region of interest, F() is the feature extraction network, z is the embedding vector representing the output of the feature extraction network, and d is the dimension of the embedding vector.

[0083] With each registered cat template vector Calculate cosine similarity:

[0084]

[0085] Find the identity corresponding to the highest similarity:

[0086]

[0087] when If the identity is confirmed, the individual is considered an unknown entity. It is the identity determination threshold parameter.

[0088] Triplet loss can be used during the training phase to enhance discriminative power.

[0089]

[0090] in, Embed for anchor samples, It's the same cat (positive). For different cats (negative), m is the interval hyperparameter.

[0091] The technical solution has been further optimized, and the door control device decides whether to allow the pet to pass based on the analysis results of the mouth area and / or the pet's status.

[0092] The strategy for gate control devices can be written as a combination of rule layers, confidence thresholds, and timing voting to ensure interpretability and engineering robustness.

[0093] If a pick-up detection frame exists and satisfies If the frame is within the ROI at the mouth, then the frame outputs "Reject"; otherwise, "Allow".

[0094] Decision based on the nearest T-frame (e.g., 3-5 frames). If the proportion of Reject (For example If the condition is true, then it will be rejected; otherwise, it will be allowed. This is the probability threshold.

[0095] If the pet is a registered pet that is allowed to pass and is not carrying anything and the pain assessment indicates no pain, then open the door;

[0096] If the object being carried is rejected, or the identity is unknown and a more conservative risk strategy is adopted, then the lockout will remain active.

[0097] If the pain level exceeds the threshold, an alarm / logging can be triggered, but it does not have to be strongly tied to access (determined by product strategy).

[0098] The door control device integrates the pet's identity recognition results, mouth area analysis results, and pain assessment results, and controls the pet door to lock or unlock according to a preset strategy.

[0099] refer to Figure 2 The present invention also provides a control method for an intelligent pet gate based on image recognition, comprising the following steps:

[0100] Images of the pet gate area are acquired using an image acquisition device;

[0101] The image is analyzed by an artificial intelligence processing unit to perform pet identification and mouth area analysis;

[0102] Based on the analysis results of the mouth area, the passage status of the pet door is controlled by the door control device.

[0103] Further optimization of the technical solution, the specific steps of oral region analysis include:

[0104] The acquired images are preprocessed;

[0105] A deep learning object detection network is used to infer the preprocessed image and output a set of detection results including candidate bounding boxes, confidence scores, and categories.

[0106] Based on the detected pet head area, the mouth area of ​​interest is determined according to a preset ratio;

[0107] Filter the detection boxes located within the region of interest in the mouth and classified as foreign objects or items;

[0108] Based on the screening results, determine whether there are any items in the mouth. If so, further classify the items.

[0109] Further optimize the technical solution. The steps for determining whether there are items at the mouth based on the screening results include: performing a time-series voting on the judgment results of multiple consecutive frames of images. Only when the proportion of judgments indicating the presence of items or the need to refuse passage reaches a preset threshold will the passage be finally determined to be refused.

[0110] Further optimization of the technical solution and control method also includes a step of analyzing the pet's state through an artificial intelligence processing unit, including pain assessment; the step of the door control device controlling the pet's passage through the door also incorporates the results of the pet's state analysis.

[0111] Application example:

[0112] YOLO 11 Training and Inference Process for Cat Picking Up Objects: An Implementation Case Study

[0113] (1) Dataset Construction

[0114] The inventor collected images / video frames of cats at the doorway, covering daytime, nighttime, backlighting, different types of objects carried, and different postures.

[0115] The label should include at least: a bounding box for the object held in the mouth / foreign object (and an optional bounding box for the cat's head / face).

[0116] The training / validation / test sets are divided into layers based on individuals and scenes to prevent consecutive frames of the same video from leaking into different sets.

[0117] (2) Model training

[0118] Input: Image I; Output: Detection box, category, and confidence score.

[0119] Using a weighted sum of detection losses:

[0120]

[0121] in, For bounding box regression loss (such as IoU / GIoU / CIoU class); For classification loss (cross entropy, etc.); Loss of confidence in the target; , and All are weighting coefficients.

[0122] (3) Reasoning and Post-processing

[0123] Remove duplicate bounding boxes using NMS (Non-maximum suppression): When the IoU between two boxes is greater than a threshold. Those with higher confidence levels are retained.

[0124] Only the bounding boxes within the mouth ROI are counted, and the output is whether an object is being held and the confidence level.

[0125] (4) Experimental results

[0126] On the test set, the accuracy rate for identifying whether an object is being carried in the mouth reached approximately 95%.

[0127] The pain recognition system was trained on FGS labeled data, and the accuracy of "whether there is pain" recognition on the test set reached about 95%.

[0128] Example of Artificial Intelligence Analysis Process

[0129] The processing flow of the artificial intelligence processing unit for a single triggered event is as follows:

[0130] Step S201: Image Acquisition and Triggering

[0131] When a pet approaches, the motion sensor triggers the image acquisition module. The module continuously captures a series of images or a short video to capture the dynamic process, combat motion blur, and provide a basis for temporal analysis.

[0132] Step S202: Image preprocessing

[0133] Each frame of the original image is preprocessed to prepare it for the standardized model input. Preprocessing may include:

[0134] Size scaling: Scaling the image to the input size required by the neural network (e.g., 640x640 pixels).

[0135] Color normalization: Normalizes pixel values ​​from [0, 255] to [0, 1] or performs mean-standard deviation normalization.

[0136] Enhancement processing (optional): Depending on the configuration, simple denoising, histogram equalization, etc., may be performed to improve image quality.

[0137] Step S203: Pet Detection and Key Region Localization. The preprocessed image is input into a deep learning object detection network (e.g., a lightweight version of YOLOv8 or YOLOv11). This network is trained to simultaneously detect two key targets: the pet's head and foreign objects / items in its mouth.

[0138] The network outputs a series of candidate detection boxes, each containing location coordinates (x, y, w, h), confidence score, and category.

[0139] To improve accuracy, a two-stage strategy can be employed. First, a model is used specifically for detecting pet heads. Then, within the detected head bounding box, the region of interest (ROI) is cropped out at a predetermined ratio (e.g., the central region of the lower half of the head bounding box). This ROI image is then fed into a second model specifically for classification. This method effectively reduces background interference and improves the accuracy of small object detection.

[0140] Step S204: Oral region analysis

[0141] This step enables the analysis of the oral region.

[0142] Presence detection: For a single frame image, check for the existence of a detection box for the category of foreign objects / items in the mouth, and ensure that its confidence level is higher than a preset threshold. If a two-stage method is used, check whether any foreign objects / items are detected within the mouth ROI.

[0143] Item Classification: For detected foreign objects / items in the mouth, the model can directly output their subcategories. The system internally maintains a mapping table that maps these subcategories to control decision categories.

[0144] To avoid false detections in a single frame leading to accidental door locking or unlocking, the system employs a multi-frame voting mechanism. The analysis results of N consecutive frames are statistically analyzed.

[0145] If the frame rate exceeds the threshold, the final decision is to deny passage.

[0146] Otherwise, the final decision is to allow passage.

[0147] This mechanism significantly improves the system's robustness in the face of instantaneous occlusion, changes in light and shadow, and adheres to the safety principle of err on the side of caution in denying access rather than allowing it to proceed.

[0148] Step S205: Pet identification (executed in parallel or serially)

[0149] The image of the pet's head region detected in step S203 is used.

[0150] The data is fed into a feature extraction network. The network outputs a high-dimensional feature vector, such as a 128-dimensional or 256-dimensional vector. This vector can compactly represent the unique appearance features of the pet's head.

[0151] The system pre-builds a feature template library for each registered pet. During registration, several head images of the pet are collected from multiple angles and under different lighting conditions. Feature vectors are extracted from each image and averaged to obtain the pet's template vector.

[0152] For the currently detected pet, calculate the cosine similarity between its feature vector and each template vector in the template library.

[0153] Find the highest similarity and its corresponding pet identity.

[0154] Determine if the highest similarity score is greater than a preset recognition threshold. If yes, identify it as a pet; otherwise, identify it as an unknown individual.

[0155] Identity information can be used for more complex control strategies, such as enabling advanced features only for pets in known households, while keeping unknown animals locked or performing only the most basic presence detection.

[0156] Step S206: Pet condition analysis, taking pain assessment as an example.

[0157] Data preparation: A large number of cat facial images were collected, and professional veterinarians were invited to rate each image according to the feline facial expression scoring criteria. This constitutes a supervised learning dataset.

[0158] Model Training: A neural network is constructed, taking a pet's facial image as input and outputting a probability distribution corresponding to different pain levels. The model is trained using a cross-entropy loss function, enabling it to learn to predict pain scores based on subtle facial features.

[0159] Online reasoning: Once the system identifies the pet, its current facial image can be input into a pre-trained pain assessment model to obtain a predicted pain level.

[0160] Decision Fusion: Pain status doesn't have to be directly used to control door locks; instead, it can serve as a health alert. For example, when the system detects that a pet has repeatedly shown high pain scores, it can send a notification to the owner via a mobile app. In more complex strategies, it can also be incorporated into access decisions.

[0161] Step S207: Comprehensive Decision Making and Gated Execution

[0162] The door control device receives information from steps S204 and S206. The system executes preset control strategy logic, for example:

[0163] Basic strategy: If the final decision of the oral analysis is REJECT, the door lock will remain locked regardless of the identity.

[0164] Authentication strategy: If the mouth analysis is ALLOW but the identity is an unknown individual, the door lock will be locked; if it is a registered pet, the door will be unlocked.

[0165] Health care strategy: If the registered pet's oral analysis is set to ALLOW, but the pain assessment exceeds the high threshold, a reminder that the pet may be uncomfortable will be sent to the owner's mobile phone at the same time as unlocking the device.

[0166] Finally, the generated control commands are sent to the control execution module to open and close the door.

[0167] 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 smart pet gate based on image recognition, characterized in that, include: Image acquisition device for acquiring images of the pet gate area; The artificial intelligence processing unit uses deep learning algorithms and / or machine learning algorithms to analyze images and performs the following functions: (a) Pet identification: Identifying individual pets by analyzing their physical characteristics; (b) Mouth area analysis: Detection and / or analysis of the pet's mouth area, which includes any one or more of the following: detecting the presence of foreign objects or items in the mouth area, identifying or classifying foreign objects or items in the mouth area, and analyzing changes in the morphology or state of the mouth area; The door control device controls the passage status of the pet door based on the analysis results of the artificial intelligence processing unit.

2. The intelligent pet gate based on image recognition according to claim 1, characterized in that, The identification or classification function in the mouth area analysis will distinguish the detected foreign objects or items into at least two categories: those that should be refused passage and those that are allowed passage. The artificial intelligence processing unit will send corresponding control commands to the door control device based on the classification results.

3. The intelligent pet gate based on image recognition according to claim 1, characterized in that, The deep learning algorithm can be any type of neural network algorithm, including convolutional neural networks, recurrent neural networks, or transformer networks; the machine learning algorithm can be any type of pattern recognition algorithm or classification algorithm.

4. The intelligent pet gate based on image recognition according to claim 1, characterized in that, The artificial intelligence processing unit also performs pet status analysis functions, which include any one or more of the following: assessing pain status by analyzing pet facial features, assessing health status by analyzing appearance features, and assessing emotional or mental state by analyzing behavioral features.

5. The intelligent pet gate based on image recognition and its control method according to claim 1, characterized in that, Mouth region analysis can employ any type of image analysis technique, including one or more of object detection, image segmentation, feature extraction, morphological analysis, anomaly detection, or semantic analysis.

6. The intelligent pet gate based on image recognition and its control method according to claim 1, characterized in that, Pet personal identification is achieved by analyzing the pet's physical characteristics, which include any one or more of facial features, body shape, coat color and pattern, body markings, or other identifiable features.

7. A control method for an intelligent pet gate based on image recognition, characterized in that, Includes the following steps: Images of the pet gate area are acquired using an image acquisition device; The image is analyzed by an artificial intelligence processing unit to perform pet identification and mouth area analysis; Based on the analysis results of the mouth area, the passage status of the pet door is controlled by the door control device.

8. The control method for an intelligent pet gate based on image recognition according to claim 7, characterized in that, The specific steps of oral region analysis include: The acquired images are preprocessed; A deep learning object detection network is used to infer the preprocessed image and output a set of detection results including candidate bounding boxes, confidence scores, and categories. Based on the detected pet head area, the mouth area of ​​interest is determined according to a preset ratio; Filter the detection boxes located within the region of interest in the mouth and classified as foreign objects or items; Based on the screening results, determine whether there are any items in the mouth. If so, further classify the items.

9. The control method for an intelligent pet gate based on image recognition according to claim 8, characterized in that, The steps for determining whether there are items in the mouth based on the screening results include: performing a time-series voting on the judgment results of multiple consecutive frames of images, and only when the proportion of judgments indicating the presence of items or the need to refuse passage reaches a preset threshold, is the final judgment to refuse passage made.

10. The control method for an intelligent pet gate based on image recognition according to claim 9, characterized in that, The control method also includes a step of analyzing the pet's state through an artificial intelligence processing unit, including a pain assessment; the step of the door control device controlling the pet's passage through the door also incorporates the results of the pet's state analysis.

Images