Pet individual recognition and personalized feeding plan generation method based on AI vision

Abstract

The application relates to the technical field of intelligent pet devices and machine vision applications, and provides a pet individual recognition and personalized feeding plan generation method and system based on AI vision. The method is applied to an intelligent pet feeder, the intelligent pet feeder is provided with a vision acquisition module; the method comprises the following steps: acquiring an appearance image of a pet to be recognized through the vision acquisition module; after standardization pretreatment is performed on the appearance image, the appearance image is input into a deep vision feature extraction network which is pre-trained and constructed based on high-dimensional feature vector extraction, core appearance features of the pet are extracted; based on the core appearance features, the species recognition and individual unique identity confirmation of the pet are completed; individual data matched with the confirmed pet individual is acquired; a personalized feeding plan is generated by combining the recognized pet species, the individual identity and the retrieved individual data; and the feeder is controlled to complete feeding according to the generated personalized feeding plan.

Description

Technical Field

[0001] This application relates to the technical field, and in particular to a method for pet individual identification and personalized feeding plan generation based on AI vision. Background Technology

[0002] With the rapid development of the Internet of Things and the pet economy, smart pet feeders have become one of the core devices for family pet ownership. Traditional timed and quantitative feeders in existing technology can only perform indiscriminate feeding according to preset fixed times and amounts, failing to identify the breed and individual identity of pets, and unable to adapt to the personalized feeding needs of different pets. In family settings with multiple pets, problems such as food competition, overfeeding, and insufficient feeding are prone to occur, especially failing to meet the special feeding needs of young, senior, and sick pets, demonstrating a serious lack of precise feeding capabilities. Summary of the Invention

[0003] This application provides a method and system for pet individual recognition and personalized feeding plan generation based on AI vision. It aims to solve the problem that some smart feeders equipped with recognition functions, although they introduce recognition technology to try to solve the above problems, most solutions use conventional convolutional neural networks to realize pet recognition. Such models have a huge amount of computation and parameters, and cannot complete real-time inference on the local embedded hardware platform of the feeder. They must rely on cloud servers for data processing and recognition calculations. Once there are network fluctuations, network outages, etc., the device cannot realize the recognition and feeding functions, resulting in extremely poor operation stability. In addition, cloud inference has significant delays, which cannot meet the needs of real-time feeding scenarios.

[0004] In a first aspect, embodiments of this application provide a method for pet individual identification and personalized feeding plan generation based on AI vision, applied to a smart pet feeder, wherein the smart pet feeder is equipped with a visual acquisition module; the method includes: The visual acquisition module acquires an image of the pet to be identified; after standardizing the image, it is input into a pre-trained deep visual feature extraction network based on high-dimensional feature vector extraction to extract the core appearance features of the pet; the core appearance features include at least fur color and body shape. Based on the core appearance features, the pet species is identified and its unique identity is confirmed; individual data matching the confirmed pet individual is obtained; the individual data includes at least age, weight, and health-related data; By combining the identified pet species, individual identity, and retrieved individual data, a personalized feeding plan is generated, which includes feeding time, single feeding amount, and suitable food type; the feeder is controlled to complete the feeding according to the generated personalized feeding plan.

[0005] In some embodiments, acquiring the appearance image of the pet to be identified through the visual acquisition module includes: real-time detection of the preset feeding area of ​​the feeder; when a pet is detected entering the area, activating the visual acquisition module to continuously acquire multiple frames of the appearance image of the pet to be identified at a preset frame rate; verifying the clarity and occlusion of the acquired multiple frames of images; and filtering out appearance images that are unobstructed and whose clarity meets a preset threshold.

[0006] In some embodiments, the method further includes: uniformly adjusting the acquired appearance image to a fixed input size preset by the deep learning model, performing illumination correction and background interference removal processing on the adjusted appearance image, and then normalizing the pixel values ​​of the appearance image to complete the standardized preprocessing of the appearance image.

[0007] In some embodiments, the method further includes: using high-dimensional feature vector extraction as the core feature extraction unit, decomposing the standard convolution operation into depthwise convolution that extracts spatial features for each input channel individually, and pointwise convolution that fuses cross-channel information for the output features of the depthwise convolution, building the main network by stacking multiple high-dimensional feature vector extraction modules, and adjusting the number of channels and input resolution of the network by preset width multipliers and resolution multipliers, thereby completing the construction of a deep visual feature extraction network based on high-dimensional feature vector extraction to adapt to the embedded hardware platform of the feeder.

[0008] In some embodiments, the step of identifying the pet species and confirming its unique identity based on the core appearance features includes: inputting the extracted core appearance features of the pet into a high-dimensional feature vector to extract and construct the species classification branch of a deep visual feature extraction network, and outputting the species identification result corresponding to the pet; converting the core appearance features into a fixed-dimensional feature vector, comparing the feature vector with the features in a pre-stored pet individual feature database, determining the corresponding pet individual with the highest similarity and meeting a preset similarity threshold, and completing the unique identity confirmation of the current pet.

[0009] In some embodiments, obtaining individual data matching the confirmed pet individual includes: after completing the unique identity confirmation of the pet individual, initiating a data retrieval request for the corresponding individual to a preset pet data management platform, obtaining the individual data of the pet's age, real-time weight, vaccination records, past medical history and dietary restrictions, and verifying the validity and completeness of the retrieved individual data.

[0010] In some embodiments, generating a personalized feeding plan that includes feeding time, single feeding amount, and suitable food type by combining the identified pet species, individual identity, and retrieved individual data includes: determining the basic feeding rules for the corresponding species based on the identified pet species; adjusting the feeding frequency, single feeding amount, and suitable food type in the basic feeding rules based on the pet individual's age, weight, and health-related data; and optimizing the feeding time nodes based on the pet's historical eating data to generate a personalized feeding plan that is suitable for the pet individual.

[0011] In some embodiments, controlling the feeder to complete feeding according to the generated personalized feeding plan includes: according to the generated personalized feeding plan, when the preset feeding time node is reached, driving the feeding execution mechanism of the feeder to start the feeding action, collecting the food output data during the feeding process, and stopping the operation of the feeding execution mechanism when the cumulative output reaches the preset threshold corresponding to the single feeding amount of this plan, thus completing a single precise feeding.

[0012] In some embodiments, the method further includes: after a single feeding is completed, acquiring an image of the remaining food in the feeding area through a visual acquisition module, identifying and calculating the actual intake of the pet during this feeding, updating the pet's individual data by combining the feeding time and the pet's eating behavior data, and dynamically optimizing and adjusting the subsequent personalized feeding plan based on the updated individual data.

[0013] In some embodiments, the method further includes: when multiple pets are detected simultaneously in the preset feeding area of ​​the feeder by the visual acquisition module, the species and individual identity of each pet are identified and confirmed respectively, the personalized feeding plan and feeding permission of each pet are retrieved, and the corresponding feeding action is executed only when the target pet that conforms to the current feeding plan is in the feeding area, and the eating behavior of non-target pets is blocked.

[0014] The method provided by this invention can complete the entire process from pet image acquisition, feature extraction, individual recognition to feeding plan generation and feeding execution on the local embedded hardware platform of the smart pet feeder. It does not rely on cloud servers for inference calculations, which completely solves the problems of poor device operation stability and insufficient real-time performance caused by the reliance on the cloud in the existing technology, and greatly improves the environmental adaptability and operational reliability of the device.

[0015] By constructing a deep visual feature extraction network based on high-dimensional feature vector extraction, the model significantly reduces the amount of computation and parameters, adapts to the limited computing power of embedded devices, and ensures the accuracy of pet feature extraction. It can simultaneously achieve accurate identification of pet species and individual unique identity, effectively solving the pain points of insufficient local recognition accuracy and inability to distinguish between different pets of the same breed in existing technologies. It also eliminates the problem of misfeeding caused by identification errors and is perfectly adapted to the feeding scenarios of multi-pet households.

[0016] By deeply integrating the pet's individual identification results with its age, weight, and health-related data, a personalized feeding plan is automatically generated for each pet, including feeding time, single feeding amount, and suitable food type. The feeding actions are precisely controlled according to the plan, realizing a complete closed loop from accurate pet identification to scientific personalized feeding. This truly meets the refined feeding needs of different pets at different physiological stages and in different health states, effectively ensuring the pet's dietary health.

[0017] This invention can be implemented directly through the embedded controller of the feeder without the need for additional high-end computing hardware. While improving device performance, it effectively controls the production cost of the device, making it easy to promote and apply on a large scale.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic flowchart illustrating the steps of a method for pet individual identification and personalized feeding plan generation based on AI vision, provided in one embodiment of this application. Figure 2 This is a schematic diagram of the structure of an intelligent pet feeder provided in one embodiment of this application; Figure 3 This is a schematic block diagram of a pet individual identification and personalized feeding plan generation system based on AI vision, provided in one embodiment of this application; Figure 4 This is a schematic block diagram of the structure of an intelligent pet feeder provided in one embodiment of this application.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0024] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0025] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0026] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0027] With the rapid development of the Internet of Things and the pet economy, smart pet feeders have become one of the core devices for family pet ownership. Traditional timed and quantitative feeders in existing technology can only perform indiscriminate feeding according to preset fixed times and amounts, failing to identify the breed and individual identity of pets, and unable to adapt to the personalized feeding needs of different pets. In family settings with multiple pets, problems such as food competition, overfeeding, and insufficient feeding are prone to occur, especially failing to meet the special feeding needs of young, senior, and sick pets, demonstrating a serious lack of precise feeding capabilities.

[0028] While some smart pet feeders equipped with recognition functions attempt to solve the above problems by introducing recognition technology, most solutions use conventional convolutional neural networks to achieve pet recognition. Such models have a huge amount of computation and parameters, making it impossible to complete real-time inference on the feeder's local embedded hardware platform. They must rely on cloud servers for data processing and recognition calculations. Once there are network fluctuations or network outages, the device cannot perform recognition and feeding functions, resulting in extremely poor operational stability. Furthermore, cloud inference has significant latency, which cannot meet the needs of real-time feeding scenarios.

[0029] To solve the above problem, please refer to Figure 1 This application provides a method for pet individual identification and personalized feeding plan generation based on AI vision, applicable to, for example... Figure 2 The smart pet feeder shown is equipped with a visual acquisition module 10. It should also be noted that all information involved in the method provided in this application was extracted with the authorization of the relevant user and in accordance with relevant regulations, and will not infringe on user privacy.

[0030] The provided AI vision-based method for pet individual identification and personalized feeding plan generation includes steps S101 to S103. Details are as follows: Step S101. Obtain the appearance image of the pet to be identified through the visual acquisition module; after standardizing the appearance image, input it into the pre-trained deep visual feature extraction network constructed based on high-dimensional feature vector extraction to extract the core appearance features of the pet; the core appearance features include at least fur color and body shape.

[0031] Specifically, this step is the foundation of the entire method. Its core objective is to acquire pet images that meet the recognition requirements and, through a pre-trained deep visual feature extraction network, extract core appearance features that can be used for pet breed differentiation and individual identification. These features include: 1. Pet appearance image acquisition is performed using a vision acquisition module integrated into the smart pet feeder. This module employs a digital camera adapted to an embedded platform, fixedly mounted directly above the feeder's outlet. The camera's field of view is vertically downward, covering the feed bowl and a pre-defined feeding area within a 30cm radius. It can clearly capture the complete torso and facial features of the pet entering the feeding area. After the device is powered on, the vision acquisition module enters a low-power standby detection state, continuously monitoring changes in the pre-defined feeding area. When a live pet is detected entering the area, the module immediately wakes up and enters normal operating mode, acquiring an RGB color image of the pet to be identified, providing raw data for subsequent processing.

[0032] 2. Standardization preprocessing of appearance images: The core objective is to convert the original acquired appearance images into standard images that conform to the input specifications of deep learning models, while eliminating the impact of environmental interference on recognition accuracy. The specific implementation process is as follows: First, the original images are resized, adjusting images of arbitrary resolution to the fixed input size preset during model training. During the adjustment process, the proportion of the pet subject remains unchanged to avoid image distortion that leads to feature loss. Then, environmental interference correction is performed on the resized images, including illumination uniformity correction for low-light and backlight scenes, and interference removal for fixed backgrounds, stripping fixed background content such as the feeder body, walls, and ground, retaining only the foreground of the pet subject. Finally, the pixel values ​​of the images are normalized, adjusting the pixel values ​​to the corresponding range during model training. This completes the entire preprocessing process, resulting in a standard image that can be directly input into the model.

[0033] 3. Core appearance feature extraction based on a deep visual feature extraction network. The core execution vehicle is a pre-trained deep visual feature extraction network constructed based on high-dimensional feature vector extraction. This model is specifically designed for the embedded hardware platform of the feeder and can complete real-time feature extraction and inference with limited computing power. After the pre-processed standard image is input into the model, it will pass through the model's multi-layer feature extraction network in sequence. Through the high-dimensional feature vector extraction structure, the low-level basic features such as edges and textures of the pet image are first extracted, and then gradually combined to form high-level core features such as the pet's coat color distribution, body contour, and body proportions. Finally, the core appearance features that can uniquely represent the pet's appearance attributes are output, which at least include body shape features that can distinguish the pet species and coat color features that can distinguish different individuals of the same species, providing core data support for subsequent species identification and individual identity confirmation.

[0034] It should also be noted that the deep visual feature extraction network can be various architectures such as ViT, CNN, MobileNet, etc., and is also compatible with possible future edge solutions. This application embodiment does not limit this.

[0035] Step S102. Based on the core appearance features, complete the pet species identification and individual unique identity confirmation; obtain individual data matching the confirmed pet individual; the individual data includes at least age, weight and health-related data.

[0036] Specifically, this step is the core and crucial link in the entire method. Its core objective is to accurately match the pet's species and individual identity based on the extracted key appearance features, while simultaneously retrieving all valid data corresponding to that individual. This provides the data basis for generating a personalized feeding plan, specifically including: 1. Pet species identification and unique individual identity verification: Based on the core appearance features extracted in step S101, two identification tasks are completed simultaneously through the dual-branch structure of a deep learning model. The specific implementation process is as follows: First, the core appearance features are input into the species classification branch of the model. This branch has been specifically trained on a pet species dataset and can output the species identification result corresponding to the pet based on features such as body size and basic coat color. It can at least distinguish between the two most common pet species, cats and dogs, and can also be extended to identify other pet species such as rabbits and hamsters. Simultaneously, the core appearance features are converted into a fixed-dimensional feature vector of a preset dimension (such as 128-dimensional, 512-dimensional, or higher). This vector can uniquely represent the individual appearance attributes of the pet. The vector is compared one by one with the pet individual feature database pre-stored locally on the feeder. The matching degree of the current vector with each pre-stored pet standard feature vector in the database is calculated, and the corresponding result with the highest matching degree is selected. When the matching degree exceeds the preset identification threshold (the identification threshold can be set between 0.3 and 0.9 according to the actual scenario), the unique identity of the current pet is confirmed, and it is clear that the pet is the corresponding individual pre-registered by the user.

[0037] 2. Individual pet data acquisition: This process is initiated immediately after confirming the unique identity of each pet. The core objective is to obtain all valid individual data for that pet. The specific implementation process is as follows: First, a data retrieval request for the corresponding individual is sent to the pet data management unit of the feeder. This data management unit consists of a local storage module and a cloud synchronization module. The local storage module stores the pet's core mandatory data and can be retrieved normally even without an internet connection. The cloud synchronization module stores the pet's complete health data and can be updated and synchronized in real time while connected to the internet. The retrieved individual data includes at least the pet's age, real-time weight, and health-related data. The health-related data may include vaccination records, past medical history, dietary restrictions, and veterinary feeding instructions. After data retrieval, the validity and completeness of the acquired individual data are verified. It is confirmed that no core mandatory data is missing, the data values ​​are within a reasonable range, and the data update time meets the preset validity period requirements. Data that passes the verification is valid individual data that can be used to generate a feeding plan.

[0038] Meanwhile, it should be noted that this application allows the device to be responsible for image detection and extraction, while the cloud is responsible for processing high-dimensional feature comparison and recognition. That is, the method provided in any embodiment of this application can also be realized through the edge-cloud collaborative architecture.

[0039] Step S103. Combine the identified pet species, individual identity and retrieved individual data to generate a personalized feeding plan that includes feeding time, single feeding amount and suitable food type; control the feeder to complete the feeding according to the generated personalized feeding plan.

[0040] Specifically, this step is the final execution stage of the entire method. Its core objective is to generate a scientifically tailored, personalized feeding plan based on the pet's identification results and individual data, and to precisely control the feeder to complete the feeding action, achieving refined feeding. This includes: 1. Personalized feeding plan generation: The core of this process is to generate a unique feeding plan by combining the pet's species attributes, individual characteristics, and health status. The specific implementation process is as follows: First, based on the pet species identified in step S102, the basic feeding rules for the corresponding species are determined. These basic rules are formulated with reference to authoritative pet feeding standards and include the basic feeding frequency, feeding amount per unit body weight, and suitable basic food types for different growth stages of the corresponding species. Then, based on the pet's individual age, weight, and health-related data, the basic feeding rules are adjusted to be personalized. For example, the single feeding amount is reduced for overweight pets, the feeding frequency and food type are adjusted for young and senior pets, and the food type and feeding plan are adjusted for sick pets according to veterinary advice. At the same time, the historical eating data of the pet are referenced to optimize the feeding time nodes that are suitable for the pet's eating habits. Finally, a complete personalized feeding plan is generated. This plan includes at least three core elements: clear daily feeding time nodes, precise feeding amount per single feeding, and food types suitable for the individual pet.

[0041] 2. Precise control of feeding execution: Following the generated personalized feeding plan, the feeder's actuator is controlled to complete precise feeding. The specific implementation process is as follows: The feeder's main control unit monitors the time in real time. When the preset feeding time node in the personalized feeding plan is reached, it first confirms that the pet currently in the feeding area is the target pet corresponding to the plan. After confirmation, it sends a start command to the feeder's feeding actuator, driving the actuator to start the feeding action. During the feeding process, the feeder's discharge detection unit collects the cumulative discharge data in real time and compares the cumulative discharge with the single feeding amount of this plan. When the cumulative discharge reaches the preset threshold corresponding to the plan, it immediately sends a stop command to the feeding actuator, closes the discharge port, terminates the feeding action, and completes this precise feeding, avoiding problems such as overfeeding or insufficient feeding.

[0042] In some embodiments, acquiring the appearance image of the pet to be identified through the visual acquisition module includes: real-time detection of the preset feeding area of ​​the feeder; when a pet is detected entering the area, activating the visual acquisition module to continuously acquire multiple frames of the appearance image of the pet to be identified at a preset frame rate; verifying the clarity and occlusion of the acquired multiple frames of images; and filtering out appearance images that are unobstructed and whose clarity meets a preset threshold.

[0043] This embodiment provides a specific implementation method for the pet appearance image acquisition step S101, which can filter high-quality and effective images. It is suitable for home use scenarios where pet postures are constantly changing and there is a lot of environmental interference, and can effectively avoid recognition errors caused by low-quality images.

[0044] After the smart pet feeder is powered on, the visual acquisition module continuously monitors the feeder's preset feeding area in real time in low-power mode. It generates bounding boxes using a single-stage object detection network, performs image segmentation, and judges changes within the area. When a live pet is detected entering the preset feeding area, the visual acquisition module is immediately activated, continuously acquiring 10 frames of the pet's appearance at a preset frame rate of 10 frames per second. After acquiring multiple frames, each frame is individually checked for sharpness and occlusion. Sharpness is checked by calculating the uniformity of grayscale distribution; higher uniformity indicates higher image sharpness. The preset sharpness threshold is a grayscale uniformity greater than 50; images below this threshold are considered blurry and invalid. Image occlusion verification is achieved by detecting the integrity of the core feature regions of the pet's face and torso. Images with a core feature region integrity greater than 80% are considered unoccluded images, while images with an integrity lower than this threshold are considered occluded and invalid images. After verification, valid appearance images that simultaneously meet the requirements of being unoccluded and having a clarity that meets the preset threshold are selected from the 10 collected images. The frame with the highest clarity is selected as the input image for subsequent preprocessing and feature extraction. If none of the collected images meet the valid criteria, a new round of multi-frame images is collected until a valid image that meets the requirements is selected.

[0045] At the same time, based on the bounding box output by the object detection model, the bounding box range is expanded outward according to a preset expansion ratio to ensure that the cropped individual pet image contains complete pet appearance features and avoid feature loss caused by boundary truncation.

[0046] In some embodiments, the method further includes: uniformly adjusting the acquired appearance image to a fixed input size preset by the deep learning model, performing illumination correction and background interference removal processing on the adjusted appearance image, and then normalizing the pixel values ​​of the appearance image to complete the standardized preprocessing of the appearance image.

[0047] This embodiment provides a fully implementable preprocessing method for the standardized preprocessing of appearance images in step S101. It can effectively eliminate environmental interference, ensure that the images of the input model meet the specifications, and improve the model recognition accuracy.

[0048] After selecting valid appearance images, the images are first resized uniformly using bilinear interpolation to adjust them to a fixed input size preset by the deep learning model, such as 224×224 or 384×384 pixels. During scaling, the aspect ratio of the pet's main subject is locked, and edge areas are filled to prevent stretching and distortion. Then, the resized images undergo illumination correction. Automatic white balance is used to correct color cast, and gamma correction is used to adjust brightness uniformity, resolving overexposure or underexposure issues in backlit or low-light environments. Background interference removal is performed using object detection. The current image is compared with a pre-stored image of the feeding area with empty background to remove fixed background content such as the feeder, ground, and walls, retaining only the foreground image of the pet, thus reducing interference from irrelevant backgrounds on feature extraction. Finally, pixel values ​​are standardized according to the normalization parameters corresponding to the model training stage, linearly converting the original pixel values ​​in the 0 to 255 range to values ​​in the 0 to 1 range, maintaining complete consistency with the preprocessing method in the model training stage. This completes the full-process standardization preprocessing, resulting in a standard image that can be directly input into a deep learning model.

[0049] In some embodiments, the method further includes: using high-dimensional feature vector extraction as the core feature extraction unit, decomposing the standard convolution operation into depthwise convolution that extracts spatial features for each input channel individually, and pointwise convolution that fuses cross-channel information for the output features of the depthwise convolution. The main network is built by stacking multiple high-dimensional feature vector extraction modules, and the number of channels and input resolution of the network are adjusted by preset width multipliers and resolution multipliers (this method is a concept unique to MobileNet and is not applicable to ViT), thus completing the construction of a deep visual feature extraction network based on high-dimensional feature vector extraction to adapt to the embedded hardware platform of the feeder.

[0050] This embodiment provides a model building and pre-training implementation method that is adaptable to the embedded hardware platform of a feeder for the deep visual feature extraction network constructed based on high-dimensional feature vector extraction in step S101, which can achieve high-accuracy feature extraction and recognition under limited computing power.

[0051] In the model building phase, high-dimensional feature vector extraction is used as the core feature extraction unit. The traditional standard convolution operation is decomposed into two independent convolution stages. The first stage is depthwise convolution, which uses a 3×3 convolution kernel for spatial feature extraction on each input channel. Each kernel corresponds to only one input channel, without cross-channel information fusion, extracting only the spatial texture and contour features of the pet image. The second stage is pointwise convolution, which uses a 1×1 convolution kernel to fuse the output of the depthwise convolution across channels, adjusting the number of output feature channels to complete a full high-dimensional feature vector extraction operation. The main network of the model is a deep network composed of multiple stacked feature extraction layers, such as one standard convolutional layer paired with 13... The high-dimensional feature vector extraction modules are stacked and constructed. At the same time, the number of channels in each layer of the network is uniformly adjusted by a preset width multiplier, and the resolution of the input image of the model is adjusted by a preset resolution multiplier. This balances the recognition accuracy and computational load of the model, making the model adaptable to the computing power limitations of the embedded hardware platform of the feeder. In the model pre-training stage, the basic feature extraction capability is first pre-trained on a large general image dataset. Then, transfer learning is used to fine-tune the model on a pet-specific image dataset. The bottom feature extraction layer of the model is frozen, and only the classification branch and feature output branch of the top layer of the model are trained. This reduces the training data requirements while improving the model's accuracy in extracting and recognizing pet features. Finally, a pre-trained model that can be directly deployed on the feeder embedded platform is obtained.

[0052] Meanwhile, the system pre-stores multiple feature vectors for each pet registered by the user in the pet individual feature database. These multiple feature vectors are composed of features extracted from multiple images of the pet taken at different times and in different postures, and stored independently. During the comparison process, the similarity between the current feature vector and the feature vectors of the corresponding pet in the database is calculated, and the maximum similarity is taken as the matching score. When the matching score exceeds the preset recognition threshold, the identity is confirmed. At the same time, when the matching confidence exceeds the preset update threshold, the current feature vector is automatically appended to the feature database of the corresponding pet, realizing the continuous dynamic update and expansion of the feature database and improving the recognition robustness in long-term use scenarios.

[0053] In some embodiments, the step of identifying the pet species and confirming its unique identity based on the core appearance features includes: inputting the extracted core appearance features of the pet into a high-dimensional feature vector to extract and construct the species classification branch of a deep visual feature extraction network, and outputting the species identification result corresponding to the pet; converting the core appearance features into a fixed-dimensional feature vector, comparing the feature vector with the features in a pre-stored pet individual feature database, determining the corresponding pet individual with the highest similarity and meeting a preset similarity threshold, and completing the unique identity confirmation of the current pet.

[0054] This embodiment provides a precise identification method with two parallel branches for the pet species identification and individual unique identity confirmation steps in step S102. It can simultaneously complete pet species classification and precise matching of individual identities within the same species, making it suitable for multi-pet households.

[0055] The core pet appearance features extracted in step S101 are simultaneously input into two parallel execution branches of the deep learning model to complete species identification and individual identity verification, respectively. The species classification branch is specifically trained on a dataset of pet images from multiple species and breeds, including cats and dogs. Based on features such as body shape and basic outline in the core appearance features, it outputs the pet's species and breed identification results. The species identification results include at least cats and dogs, and the breed identification results cover common pet cat and dog breeds. The individual identity verification branch first converts the extracted core appearance features into a 128-dimensional fixed-dimensional feature vector, which uniquely represents the pet. The feeder collects individual attributes such as fur color distribution and local features, and then compares this feature vector with a pre-stored pet individual feature database. This database stores the standard feature vectors of each pet registered by the user. The standard feature vectors are generated by averaging the features extracted from multi-scene, multi-angle images of the pet collected during the user's initial registration. During the comparison process, the matching degree between the current feature vector and each standard feature vector in the database is calculated, and the pet with the highest matching degree is selected. When the highest matching degree is greater than the preset 90% qualified threshold, the unique identity of the current pet is confirmed. If the highest matching degree is lower than the qualified threshold, the pet is determined to be an unregistered pet, and a prompt message is sent to the user.

[0056] In some embodiments, obtaining individual data matching the confirmed pet individual includes: after completing the unique identity confirmation of the pet individual, initiating a data retrieval request for the corresponding individual to a preset pet data management platform, obtaining the individual data of the pet's age, real-time weight, vaccination records, past medical history and dietary restrictions, and verifying the validity and completeness of the retrieved individual data.

[0057] This embodiment provides a highly reliable data retrieval and verification method that combines local and cloud computing for the pet individual data acquisition stage in step S102. It can ensure the effective retrieval of core data in both offline and online environments, providing accurate data support for the generation of feeding plans.

[0058] After confirming the unique identity of each pet, the feeder's main control unit first sends a request to the local storage module to retrieve the pet's core data. The local storage module pre-stores essential data for each registered pet, including the pet's age, most recently updated weight, and key dietary restrictions. This data is cached locally with each cloud update and can be retrieved even without a network connection. After retrieving the local core data, the main control unit checks the device's network connection. If the device is connected to the network, it sends a request to the pre-configured cloud-based pet data management platform to retrieve the pet's full data, including its real-time updated weight and complete vaccination records. The system collects complete individual data, including past medical history, doctor's orders, feeding requirements, and dietary restrictions, and updates the locally cached core data with the latest data obtained from the cloud. After data retrieval, the system verifies the validity and completeness of the acquired individual data. First, it verifies whether the core required fields such as age and weight are complete and without missing information. Then, it verifies whether the data values ​​are within a reasonable range. For example, the pet's age should not exceed the normal lifespan range for the corresponding species, and the weight data should conform to the reasonable weight range for the corresponding breed. Finally, it verifies the data update time. The update time of the weight data should not exceed the preset 7-day validity period. If the validity period is exceeded, the user is prompted to update the data. The individual data that passes the verification is valid data that can be used to generate a feeding plan.

[0059] In some embodiments, generating a personalized feeding plan that includes feeding time, single feeding amount, and suitable food type by combining the identified pet species, individual identity, and retrieved individual data includes: determining the basic feeding rules for the corresponding species based on the identified pet species; adjusting the feeding frequency, single feeding amount, and suitable food type in the basic feeding rules based on the pet individual's age, weight, and health-related data; and optimizing the feeding time nodes based on the pet's historical eating data to generate a personalized feeding plan that is suitable for the pet individual.

[0060] This embodiment provides a scientifically adapted feeding plan generation method for the personalized feeding plan generation step S103. It can generate a personalized feeding plan that is completely adapted to the individual pet based on the pet's species, growth stage, and health status, thereby achieving refined and scientific feeding.

[0061] First, based on the pet type and breed identified in step S102, the preset basic feeding rules for the corresponding species and breed are retrieved. These basic feeding rules are formulated with reference to pet feeding standards published by authoritative animal husbandry and veterinary institutions, clearly defining the basic feeding frequency, daily food intake per unit body weight, and suitable basic food types for the corresponding species at three different growth stages: juvenile, adult, and senior. For example, the basic rule for a healthy adult domestic cat is a daily feeding intake of 25g of adult cat food per kilogram of body weight, divided into two feedings per day, using complete adult cat food. Then, based on the pet's individual age, weight, and health-related data, the basic feeding rules are personalized. The adjustments include: for pets exceeding the standard weight for the corresponding breed, the daily food intake is reduced by 10% to 20%; for pets below the standard weight... For pets with pre-existing conditions, increase the daily food intake by 10% to 20%. For young and senior pets, increase the daily feeding frequency and reduce the amount of food per feeding, while switching to a special food suitable for the corresponding growth stage. For pets with a history of illness, dietary restrictions, or veterinary feeding requirements, adjust the food type, amount, and frequency according to the veterinarian's instructions. Then, retrieve the pet's historical eating data, analyze the pet's daily eating habits, and adjust the feeding times to the high-frequency periods when the pet actively eats. Finally, integrate all the adjusted parameters to generate a complete personalized feeding plan, clearly marking the daily feeding times, the amount of food per feeding at each time point, and the type of food suitable for the pet. This plan can be manually modified by the user, and the modified plan will be updated synchronously in the pet's individual data.

[0062] In some embodiments, controlling the feeder to complete feeding according to the generated personalized feeding plan includes: according to the generated personalized feeding plan, when the preset feeding time node is reached, driving the feeding execution mechanism of the feeder to start the feeding action, collecting the food output data during the feeding process, and stopping the operation of the feeding execution mechanism when the cumulative output reaches the preset threshold corresponding to the single feeding amount of this plan, thus completing a single precise feeding.

[0063] This embodiment provides a closed-loop precision feeding control implementation method for the feeding execution control link in step S103, which can monitor the feeding amount in real time, avoid overfeeding or underfeeding, and ensure the accurate implementation of the feeding plan.

[0064] The feeding mechanism of the smart pet feeder uses a stepper motor paired with a reduction gear set and a discharge impeller. The discharge volume is pre-calibrated to determine the corresponding discharge weight for each fixed step of the stepper motor, forming calibration parameters that correspond to step length and discharge volume, which are pre-stored in the main control unit. The main control unit monitors the system time in real time. When the preset feeding time node in the personalized feeding plan is reached, it first uses a visual acquisition module to confirm that the pet in the current preset feeding area is the target pet for the plan. After confirmation, based on the single feeding amount of this plan and the pre-stored calibration parameters, it calculates the total step length that the stepper motor needs to rotate. Then, the main control unit... A pulse signal is sent to the stepper motor's driver board, driving the stepper motor to rotate the discharge impeller and initiating the feeding action. During feeding, a high-precision weight sensor installed under the feeding bowl collects real-time data on the weight of the food inside the bowl, updating the cumulative discharge weight every second. The real-time cumulative discharge weight is compared with the planned single feeding amount. When the cumulative discharge weight reaches the preset threshold corresponding to the plan, the main control unit immediately stops sending pulse signals, the stepper motor stops rotating, the discharge port is locked, and the feeding action is terminated. After feeding is completed, the actual feeding amount is recorded and updated to the pet's historical feeding data, completing this precise feeding.

[0065] In some embodiments, the method further includes: after a single feeding is completed, acquiring an image of the remaining food in the feeding area through a visual acquisition module, identifying and calculating the actual intake of the pet during this feeding, updating the pet's individual data by combining the feeding time and the pet's eating behavior data, and dynamically optimizing and adjusting the subsequent personalized feeding plan based on the updated individual data.

[0066] This embodiment provides a supplementary implementation method for closed-loop optimization after feeding, which can dynamically optimize and adjust the subsequent personalized feeding plan based on the pet's actual eating data, continuously improve the adaptability of the feeding plan, and meet the pet's actual eating needs.

[0067] After the single feeding action in step S103 is completed, wait for the preset 30-minute feeding time. After the pet finishes eating and leaves the feeding area, activate the visual acquisition module to capture an image of the remaining food in the feeding bowl. Compare this image with the image of the empty bowl before feeding and the image of the full bowl after feeding. Calculate the volume and weight of the remaining food in the bowl through image recognition. Subtract the weight of the remaining food from the planned single feeding amount to obtain the pet's actual intake for this feeding. At the same time, the visual acquisition module records the feeding time, i.e., the total time from the start of the feeding action to the pet's final departure from the feeding area, as well as the pet's eating behavior data, including whether it is picky about food, whether there is any leftover of a certain type of food, and the eating speed. The actual intake for this feeding is then recorded. The feeding time and feeding behavior data are synchronously updated to the pet's individual database and historical feeding data. Based on the updated feeding data, the main control unit dynamically optimizes and adjusts the subsequent personalized feeding plan. The adjustment rules include: if the pet's actual intake is less than 70% of the feeding amount for three consecutive times, the amount of food for each subsequent feeding will be reduced by 10%; if the pet finishes the feeding amount for two consecutive times and stays in the feeding area to forage for food after eating, the amount of food for each subsequent feeding will be increased by 10%; if the pet consistently leaves a certain type of food uneaten, the appropriate food type in the feeding plan will be adjusted; if the pet eats too quickly, the feeding will be adjusted to a mode of feeding small amounts multiple times. Ultimately, the feeding plan is continuously and dynamically optimized to meet the pet's actual eating needs.

[0068] In some embodiments, the method further includes: when multiple pets are detected simultaneously in the preset feeding area of ​​the feeder by the visual acquisition module, the species and individual identity of each pet are identified and confirmed respectively, the personalized feeding plan and feeding permission of each pet are retrieved, and the corresponding feeding action is executed only when the target pet that conforms to the current feeding plan is in the feeding area, and the eating behavior of non-target pets is blocked.

[0069] This embodiment provides an implementation method for managing feeding permissions in a multi-pet scenario, which can effectively solve the problems of food fighting and accidental feeding in multi-pet households and ensure the accurate implementation of each pet's exclusive feeding plan.

[0070] In the image acquisition stage of step S101, when the visual acquisition module detects that there are two or more pets simultaneously within the preset feeding area of ​​the feeder, it immediately performs target detection and image segmentation on the multiple pets in the image, selecting an independent image area for each pet. For each independent image area, the preprocessing and core appearance feature extraction in step S101 and the species identification and individual identity confirmation in step S102 are performed respectively to obtain the individual identity information of each pet in the image. Subsequently, for each pet with confirmed identity, the corresponding personalized feeding plan and feeding permissions are retrieved to determine whether the current time is within the pet's feeding time window and whether the pet has the right to use the food during the current time period. Feeding permissions: After the judgment is completed, the main control unit will only start the corresponding feeding action and execute the feeding control process of step S103 when the target pet that meets the current feeding plan is alone in the preset feeding area. If a non-target pet is detected to be in the feeding area at the same time, the feeding action will be stopped immediately, and a preset prompt sound will be issued through the voice playback module on the feeder to drive the non-target pet away from the feeding area. The status of the pets in the feeding area will be continuously monitored until only the target pet is alone in the feeding area before the feeding action is resumed. Feeding non-target pets is prohibited throughout the process to avoid the problem of food snatching and accidental feeding, and to ensure the accurate execution of the exclusive feeding plan for each pet.

[0071] In some embodiments, for families with multiple pets, there are pets with special health conditions requiring strict dietary control (such as pets with chronic kidney disease, diabetes, or post-operative recovery) and healthy pets. It is necessary to strictly distinguish between prescription food and regular food, prevent cross-individual feeding, and monitor the feeding process of each pet in real time to dynamically adapt the feeding plan to the health management requirements. This solves the core pain points of existing technologies that can only achieve single-identity triggered feeding, cannot control the entire feeding process, cannot prevent cross-individual feeding, and cannot adapt to multi-level health management needs.

[0072] The visual acquisition module continuously monitors the preset feeding area at a frame rate of 15 frames per second. When at least one pet is detected entering the area, multi-target detection and image segmentation are immediately initiated. Each pet in the image is independently bounded and tracked, and a unique temporary tracking ID is assigned to each pet to continuously lock the position and posture of each pet. For each tracked pet, a clear appearance image is acquired every 2 frames, and standardized preprocessing is performed simultaneously. The image is then input into a pre-trained deep visual feature extraction network to extract the core appearance features of each pet's fur color and body shape in real time. At the same time, the relative position features of the pet's mouth and feeding port are extracted to determine whether the pet is in an effective feeding state.

[0073] For each pet, the core appearance features extracted in real time are continuously compared with the pre-stored pet individual feature database. A full identity verification is completed every 500 milliseconds to confirm the unique identity of each tracked pet. At the same time, the relative position of the pet and the feeding hole is locked to predict the risk of non-target pets approaching the feeding hole for cross-contamination. After the identity is confirmed, the health rating label and full individual data of the corresponding pet are retrieved. The health rating is divided into three levels: strict control level, regular control level, and healthy level. The dietary restrictions, prescription food compatibility requirements, and veterinary feeding rules of the corresponding pet are retrieved simultaneously to complete the data validity and timeliness verification.

[0074] Based on the pet's health rating and individual data, a personalized feeding plan is generated for each rating. The feeder uses a multi-compartment independent dispensing structure, with each compartment corresponding to a different type of food. Each compartment only grants dispensing access to pets matching the corresponding food type. Feeding is initiated only when the target pet is present in the feeding area, the target pet's health rating matches the feeding plan, and the target pet's mouth is within the effective feeding range of the feeding opening. During the feeding process, the pet's identity is continuously verified, with identity confirmation completed every 300 milliseconds. If a non-target pet is detected entering the feeding opening during feeding or eating, the feeder will not be activated. Within a 10-centimeter warning zone, feeding is immediately suspended, and the physical barrier at the feeding port is closed to prevent non-target pets from eating together. A voice prompt is issued to drive away the non-target pets until they leave the warning zone and only the target pet remains. Only then is the barrier reopened and feeding resumed. During the feeding process, the feeding time, chewing frequency, and actual intake of each pet are collected in real time and updated to the individual data of the corresponding pet. For pets under strict control, the amount and time of the next feeding are dynamically adjusted based on the actual intake and veterinary requirements after each feeding, forming a closed-loop health management feeding process.

[0075] In some embodiments, abnormal eating behaviors in pets (anorexia, overeating, difficulty chewing, pain while eating, pica) are early signs of oral diseases, digestive system diseases, endocrine diseases, etc. Existing technologies can only count the amount of food a pet eats, but cannot identify subtle abnormal behaviors during the eating process, cannot provide early warning of diseases, and cannot adaptively adjust the feeding plan based on abnormal behaviors, which is the core pain point that can easily lead to the aggravation of the pet's condition.

[0076] After a pet enters the feeding area, the visual acquisition module not only captures static images of the pet but also continuously acquires a sequence of temporal images of the entire feeding process at a rate of 20 frames per second. After standardization preprocessing of the temporal image sequence, it is input into a temporal feature extraction network optimized based on high-dimensional feature vector extraction. This network adds a temporal feature extraction branch to the original core appearance feature extraction. By analyzing the feature differences between consecutive frames, it extracts the pet's feeding behavior features, including mouth opening and closing frequency, chewing amplitude, head shaking frequency, feeding movement trajectory, licking frequency, and feeding pause duration. Simultaneously, it synchronizes continuous data from the feeding bowl weight sensor to extract the pet's feeding speed and changes in single feeding amount. The extracted static appearance features and temporal behavior features are fused to form a fusion feature that can simultaneously represent the pet's identity and feeding behavior status, which is then used for subsequent identity verification and abnormal behavior identification.

[0077] Based on static appearance features from the fusion features, the system completes the identification of the pet species and the confirmation of its unique identity. It retrieves the pet's historical eating behavior baseline data, age, weight, past medical history, and veterinary health data, as well as the standard baseline data of normal eating behavior for the species. The extracted temporal behavioral features are compared with the pet's own historical behavioral baseline and the species standard baseline. Through a pre-trained abnormal behavior classification model, five core abnormal behaviors are identified: chewing difficulties, anorexia, overeating, eating pain, and pica. The system verifies the persistence of the identified abnormal behaviors. If the abnormal behavior persists for more than two feeding cycles, it is determined to be a valid health abnormality signal, which is then updated to the pet's health data to generate corresponding health warning information.

[0078] If no valid abnormal behavior is identified, a standard personalized feeding plan is generated based on the existing pet type and individual data. If valid abnormal behavior is identified, the feeding plan is immediately adaptively adjusted based on the type of abnormal behavior and the pet's health data: For abnormal behaviors such as difficulty chewing or pain while eating, the food type is automatically adjusted to easily chewable types such as liquid food or wet food, and the single feeding amount is divided into 4-6 small, frequent feedings per day to reduce the pet's feeding burden; for anorexia, the feeding time is adjusted to the period when the pet is most active, and suitable highly palatable foods are added to the feeding plan, while the single feeding amount is reduced to increase the pet's willingness to eat; for overeating, the single feeding amount is divided into multiple... Intermittent feeding controls the pet's eating speed and adjusts the total daily food intake to a reasonable range to avoid overfeeding. For abnormal behaviors that persist for more than three feeding cycles, in addition to adjusting the feeding plan, abnormal behavior data and eating data are automatically pushed to the user's bound terminal and the preset veterinarian terminal to issue early disease warnings. At the same time, the feeding plan control rules can be automatically updated based on the veterinarian's feedback. After each adjusted feeding plan is executed, the pet's eating behavior data is continuously collected. If the abnormal behavior disappears and the eating behavior returns to the baseline range, the regular feeding plan is gradually restored, forming a closed-loop adaptive feeding process of "behavior recognition - anomaly judgment - plan adjustment - effect verification".

[0079] In some embodiments, for smart feeders on low-end embedded hardware platforms (such as STM32 and ESP32 series MCUs) with severely limited resources, computing power and storage space are extremely small, making it impossible to rely on cloud training. When users add new pets, there is a catastrophic problem of forgetting the original pet due to a decrease in recognition accuracy. At the same time, the feeder needs to run multiple tasks in parallel, such as recognition, motor control, sensor acquisition, and user communication, which can easily lead to the core pain points of recognition lag, feeding delay, and device crashes caused by insufficient resources.

[0080] Based on the core structure of high-dimensional feature vector extraction, the deep visual feature extraction network is optimized to the extreme. By using preset width and resolution multipliers, the number of model parameters is compressed to less than 0.5M, and the computational cost is controlled to less than 50MFLOPs, adapting to the computing power limitations of low-end MCUs, while retaining the core pet appearance feature extraction capability. The model adopts an architecture that completely separates the feature extraction layer and the classification head. The bottom feature extraction layer is frozen. This layer is pre-trained on a large pet dataset and has general pet feature extraction capabilities. Only the top individual classification head is retained for training, providing a foundation for incremental learning on the edge. Pet images acquired by the visual acquisition module are pre-processed and input into the model. The frozen feature extraction layer extracts core appearance features of fixed dimensions, which are then input into the top classification head to complete species identification and individual identity confirmation. The entire feature extraction and inference process only consumes no more than 70% of the MCU's computing power, ensuring real-time inference.

[0081] When a user needs to add a new pet, incremental learning can be completed directly on the feeder without connecting to a cloud server. The user collects multi-scene, multi-angle images of the new pet's appearance through the visual acquisition module. The feeder's main control unit extracts multiple sets of core appearance features of the new pet through a frozen feature extraction layer, averages them to generate the pet's standard feature vector, and stores it in the local pet individual feature database. During incremental learning, knowledge distillation is used to retain the feature vector distribution of the original registered pets, adding only the corresponding classification dimension for the new pet to the top-level classification header. At the same time, a small amount of old sample feature replay is used to avoid the catastrophic forgetting problem of decreased recognition accuracy of the original pet after adding a new pet. The entire incremental learning process is only executed during the feeder's idle time and does not affect normal feeding and recognition functions. After incremental learning is completed, the user can enter the age, weight, and health-related data of the new pet, which are stored in the local individual database to complete the full information registration of the new pet. Subsequently, individual recognition and personalized feeding plan generation can be performed normally.

[0082] The feeder's main control unit establishes a multi-task priority scheduling mechanism, dividing tasks into core real-time tasks and non-core background tasks. Core real-time tasks include pet image acquisition and inference recognition, feeding actuator control, and real-time feed output monitoring, with the highest priority. Non-core background tasks include user communication data synchronization, incremental learning training, historical data storage, and device status self-checks, with a low priority. When the visual acquisition module detects a pet entering the preset feeding area, the resource scheduling mechanism is immediately triggered, pausing all non-core background tasks and allocating over 90% of the MCU's computing power and all memory resources to the core real-time tasks, ensuring the real-time performance of pet recognition and inference, and feeding control. The system ensures high accuracy, eliminating issues such as recognition lag and feeding delays. When the pet leaves the feeding area and there is no feeding task to perform, the main control unit resumes the execution of non-core background tasks. During idle periods, it completes incremental learning, data synchronization, and device self-checks, while switching the visual acquisition module to a low-power standby mode, retaining only the low-frame-rate area detection function to reduce the overall power consumption of the device. Based on the individual pet recognition results and personalized feeding plan, the main control unit is woken up in advance before the feeding time arrives to complete the pre-allocation of resources for core tasks, ensuring the accurate execution of the feeding plan. Even on hardware platforms with extremely limited resources, stable and accurate pet recognition and personalized feeding functions can be achieved.

[0083] Please see Figure 3 As shown, Figure 3This is a schematic diagram of the structure of the AI ​​vision-based pet individual identification and personalized feeding plan generation system 200 provided in this application embodiment. The AI ​​vision-based pet individual identification and personalized feeding plan generation system 200 is used to execute the steps of the AI ​​vision-based pet individual identification and personalized feeding plan generation method shown in the above embodiments. The AI ​​vision-based pet individual identification and personalized feeding plan generation system 200 can be a single server or a server cluster, or it can be a terminal, such as a handheld terminal, laptop computer, wearable device, or robot.

[0084] like Figure 3 As shown, the AI ​​vision-based pet individual recognition and personalized feeding plan generation system 200 includes: Image acquisition unit 201 is used to acquire an appearance image of the pet to be identified through the visual acquisition module; after standardizing the appearance image, it is input into a pre-trained deep visual feature extraction network based on high-dimensional feature vector extraction to extract the core appearance features of the pet; the core appearance features include at least fur color and body shape. The data acquisition unit 202 is used to identify the pet species and confirm the unique identity of the individual based on the core appearance features; acquire individual data that matches the confirmed pet individual; the individual data includes at least age, weight and health-related data; The plan generation unit 203 is used to combine the identified pet species, individual identity and retrieved individual data to generate a personalized feeding plan that includes feeding time, single feeding amount and suitable food type; and to control the feeder to complete the feeding according to the generated personalized feeding plan.

[0085] In some embodiments, acquiring the appearance image of the pet to be identified through the visual acquisition module includes: real-time detection of the preset feeding area of ​​the feeder; when a pet is detected entering the area, activating the visual acquisition module to continuously acquire multiple frames of the appearance image of the pet to be identified at a preset frame rate; verifying the clarity and occlusion of the acquired multiple frames of images; and filtering out appearance images that are unobstructed and whose clarity meets a preset threshold.

[0086] In some embodiments, the method further includes: uniformly adjusting the acquired appearance image to a fixed input size preset by the deep learning model, performing illumination correction and background interference removal processing on the adjusted appearance image, and then normalizing the pixel values ​​of the appearance image to complete the standardized preprocessing of the appearance image.

[0087] In some embodiments, the method further includes: using high-dimensional feature vector extraction as the core feature extraction unit, decomposing the standard convolution operation into depthwise convolution that extracts spatial features for each input channel individually, and pointwise convolution that fuses cross-channel information for the output features of the depthwise convolution, building the main network by stacking multiple high-dimensional feature vector extraction modules, and adjusting the number of channels and input resolution of the network by preset width multipliers and resolution multipliers, thereby completing the construction of a deep visual feature extraction network based on high-dimensional feature vector extraction to adapt to the embedded hardware platform of the feeder.

[0088] In some embodiments, the step of identifying the pet species and confirming its unique identity based on the core appearance features includes: inputting the extracted core appearance features of the pet into a high-dimensional feature vector to extract and construct the species classification branch of a deep visual feature extraction network, and outputting the species identification result corresponding to the pet; converting the core appearance features into a fixed-dimensional feature vector, comparing the feature vector with the features in a pre-stored pet individual feature database, determining the corresponding pet individual with the highest similarity and meeting a preset similarity threshold, and completing the unique identity confirmation of the current pet.

[0089] In some embodiments, obtaining individual data matching the confirmed pet individual includes: after completing the unique identity confirmation of the pet individual, initiating a data retrieval request for the corresponding individual to a preset pet data management platform, obtaining the individual data of the pet's age, real-time weight, vaccination records, past medical history and dietary restrictions, and verifying the validity and completeness of the retrieved individual data.

[0090] In some embodiments, generating a personalized feeding plan that includes feeding time, single feeding amount, and suitable food type by combining the identified pet species, individual identity, and retrieved individual data includes: determining the basic feeding rules for the corresponding species based on the identified pet species; adjusting the feeding frequency, single feeding amount, and suitable food type in the basic feeding rules based on the pet individual's age, weight, and health-related data; and optimizing the feeding time nodes based on the pet's historical eating data to generate a personalized feeding plan that is suitable for the pet individual.

[0091] In some embodiments, controlling the feeder to complete feeding according to the generated personalized feeding plan includes: according to the generated personalized feeding plan, when the preset feeding time node is reached, driving the feeding execution mechanism of the feeder to start the feeding action, collecting the food output data during the feeding process, and stopping the operation of the feeding execution mechanism when the cumulative output reaches the preset threshold corresponding to the single feeding amount of this plan, thus completing a single precise feeding.

[0092] In some embodiments, the method further includes: after a single feeding is completed, acquiring an image of the remaining food in the feeding area through a visual acquisition module, identifying and calculating the actual intake of the pet during this feeding, updating the pet's individual data by combining the feeding time and the pet's eating behavior data, and dynamically optimizing and adjusting the subsequent personalized feeding plan based on the updated individual data.

[0093] In some embodiments, the method further includes: when multiple pets are detected simultaneously in the preset feeding area of ​​the feeder by the visual acquisition module, the species and individual identity of each pet are identified and confirmed respectively, the personalized feeding plan and feeding permission of each pet are retrieved, and the corresponding feeding action is executed only when the target pet that conforms to the current feeding plan is in the feeding area, and the eating behavior of non-target pets is blocked.

[0094] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the AI ​​vision-based pet individual identification and personalized feeding plan generation system and its modules described above can be found in the corresponding contents of the various embodiments of the AI ​​vision-based pet individual identification and personalized feeding plan generation method, and will not be repeated here.

[0095] The aforementioned method for AI-based vision-based pet identification and personalized feeding plan generation can be implemented as a computer program, which can be used in various ways, such as... Figure 3 It runs on the device shown.

[0096] Please see Figure 4 , Figure 4 This is a schematic block diagram of the structure of the intelligent pet feeder provided in this application embodiment. The intelligent pet feeder includes a processor, a memory, and a network interface connected via a device bus, wherein the memory may include a storage medium and internal memory.

[0097] The storage medium may store operating devices and computer programs. The computer program includes program instructions that, when executed, cause the processor to perform any AI vision-based method for pet individual identification and personalized feeding plan generation.

[0098] The processor provides computing and control capabilities to support the operation of the entire smart pet feeder.

[0099] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to execute any AI vision-based method for pet individual identification and personalized feeding plan generation.

[0100] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the terminal to which the present application is applied. A specific smart pet feeder may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0101] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0102] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: The visual acquisition module acquires an image of the pet to be identified; after standardizing the image, it is input into a pre-trained deep visual feature extraction network based on high-dimensional feature vector extraction to extract the core appearance features of the pet; the core appearance features include at least fur color and body shape. Based on the core appearance features, the pet species is identified and its unique identity is confirmed; individual data matching the confirmed pet individual is obtained; the individual data includes at least age, weight, and health-related data; By combining the identified pet species, individual identity, and retrieved individual data, a personalized feeding plan is generated, which includes feeding time, single feeding amount, and suitable food type; the feeder is controlled to complete the feeding according to the generated personalized feeding plan.

[0103] In some embodiments, acquiring the appearance image of the pet to be identified through the visual acquisition module includes: real-time detection of the preset feeding area of ​​the feeder; when a pet is detected entering the area, activating the visual acquisition module to continuously acquire multiple frames of the appearance image of the pet to be identified at a preset frame rate; verifying the clarity and occlusion of the acquired multiple frames of images; and filtering out appearance images that are unobstructed and whose clarity meets a preset threshold.

[0104] In some embodiments, the method further includes: uniformly adjusting the acquired appearance image to a fixed input size preset by the deep learning model, performing illumination correction and background interference removal processing on the adjusted appearance image, and then normalizing the pixel values ​​of the appearance image to complete the standardized preprocessing of the appearance image.

[0105] In some embodiments, the method further includes: using high-dimensional feature vector extraction as the core feature extraction unit, decomposing the standard convolution operation into depthwise convolution that extracts spatial features for each input channel individually, and pointwise convolution that fuses cross-channel information for the output features of the depthwise convolution, building the main network by stacking multiple high-dimensional feature vector extraction modules, and adjusting the number of channels and input resolution of the network by preset width multipliers and resolution multipliers, thereby completing the construction of a deep visual feature extraction network based on high-dimensional feature vector extraction to adapt to the embedded hardware platform of the feeder.

[0106] In some embodiments, the step of identifying the pet species and confirming its unique identity based on the core appearance features includes: inputting the extracted core appearance features of the pet into a high-dimensional feature vector to extract and construct the species classification branch of a deep visual feature extraction network, and outputting the species identification result corresponding to the pet; converting the core appearance features into a fixed-dimensional feature vector, comparing the feature vector with the features in a pre-stored pet individual feature database, determining the corresponding pet individual with the highest similarity and meeting a preset similarity threshold, and completing the unique identity confirmation of the current pet.

[0107] In some embodiments, obtaining individual data matching the confirmed pet individual includes: after completing the unique identity confirmation of the pet individual, initiating a data retrieval request for the corresponding individual to a preset pet data management platform, obtaining the individual data of the pet's age, real-time weight, vaccination records, past medical history and dietary restrictions, and verifying the validity and completeness of the retrieved individual data.

[0108] In some embodiments, generating a personalized feeding plan that includes feeding time, single feeding amount, and suitable food type by combining the identified pet species, individual identity, and retrieved individual data includes: determining the basic feeding rules for the corresponding species based on the identified pet species; adjusting the feeding frequency, single feeding amount, and suitable food type in the basic feeding rules based on the pet individual's age, weight, and health-related data; and optimizing the feeding time nodes based on the pet's historical eating data to generate a personalized feeding plan that is suitable for the pet individual.

[0109] In some embodiments, controlling the feeder to complete feeding according to the generated personalized feeding plan includes: according to the generated personalized feeding plan, when the preset feeding time node is reached, driving the feeding execution mechanism of the feeder to start the feeding action, collecting the food output data during the feeding process, and stopping the operation of the feeding execution mechanism when the cumulative output reaches the preset threshold corresponding to the single feeding amount of this plan, thus completing a single precise feeding.

[0110] In some embodiments, the method further includes: after a single feeding is completed, acquiring an image of the remaining food in the feeding area through a visual acquisition module, identifying and calculating the actual intake of the pet during this feeding, updating the pet's individual data by combining the feeding time and the pet's eating behavior data, and dynamically optimizing and adjusting the subsequent personalized feeding plan based on the updated individual data.

[0111] In some embodiments, the method further includes: when multiple pets are detected simultaneously in the preset feeding area of ​​the feeder by the visual acquisition module, the species and individual identity of each pet are identified and confirmed respectively, the personalized feeding plan and feeding permission of each pet are retrieved, and the corresponding feeding action is executed only when the target pet that conforms to the current feeding plan is in the feeding area, and the eating behavior of non-target pets is blocked.

[0112] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the steps of the AI ​​vision-based pet individual identification and personalized feeding plan generation method provided in any embodiment of this application.

[0113] The computer-readable storage medium can be the internal storage unit of the smart pet feeder described in the foregoing embodiments, such as the hard drive or memory of the smart pet feeder. The computer-readable storage medium can also be an external storage device of the smart pet feeder, such as a plug-in hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the smart pet feeder.

[0114] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for individual pet identification and personalized feeding plan generation based on AI vision, applied to a smart pet feeder, wherein the smart pet feeder is equipped with a visual acquisition module; characterized in that, The method includes: The visual acquisition module acquires an image of the pet to be identified; after standardizing the image, it is input into a pre-trained deep visual feature extraction network based on high-dimensional feature vector extraction to extract the core appearance features of the pet; the core appearance features include at least fur color and body shape. Based on the core appearance features, the pet species is identified and its unique identity is confirmed; individual data matching the confirmed pet individual is obtained; the individual data includes at least age, weight, and health-related data; By combining the identified pet species, individual identity, and retrieved individual data, a personalized feeding plan is generated, which includes feeding time, single feeding amount, and suitable food type; the feeder is controlled to complete the feeding according to the generated personalized feeding plan.

2. The method according to claim 1, characterized in that, The process of acquiring the appearance image of the pet to be identified through the visual acquisition module includes: The system detects the preset feeding area of ​​the feeder in real time. When a pet is detected entering the area, the visual acquisition module is activated to continuously acquire multiple frames of the pet's appearance at a preset frame rate. The acquired multiple frames are then checked for clarity and occlusion, and the appearance images that are unobstructed and whose clarity meets the preset threshold are selected.

3. The method according to claim 1, characterized in that, The method further includes: The acquired appearance images are uniformly adjusted to the fixed input size preset by the deep learning model. The adjusted appearance images are then subjected to illumination correction and background interference removal. Finally, the pixel values ​​of the appearance images are normalized to complete the standardization preprocessing of the appearance images.

4. The method according to claim 1, characterized in that, The method further includes: High-dimensional feature vector extraction is used as the core feature extraction unit. The standard convolution operation is decomposed into depthwise convolution, which extracts spatial features for each input channel individually, and pointwise convolution, which fuses cross-channel information on the output features of depthwise convolution. The main network is built by stacking multiple high-dimensional feature vector extraction modules. At the same time, the number of channels and input resolution of the network are adjusted by preset width multipliers and resolution multipliers. The construction of a deep visual feature extraction network based on high-dimensional feature vector extraction is completed to adapt to the embedded hardware platform of the feeder.

5. The method according to claim 4, characterized in that, The process of identifying the pet's species and confirming its unique identity based on the core appearance features includes: The extracted core appearance features of the pet are input into a high-dimensional feature vector to extract and construct the species classification branch of a deep visual feature extraction network, and the pet species identification result is output. The core appearance features are converted into fixed-dimensional feature vectors. The feature vectors are then compared with features in a pre-stored pet individual feature database to determine the corresponding pet individual with the highest similarity and that meets the preset similarity threshold, thus completing the unique identification of the current pet.

6. The method according to claim 1, characterized in that, The individual data matching the acquired and confirmed pet individuals includes: After confirming the unique identity of each pet, a data retrieval request is sent to the pre-set pet data management platform to obtain the pet's age, real-time weight, vaccination records, medical history, and dietary restrictions. The validity and completeness of the retrieved individual data are then verified.

7. The method according to claim 1, characterized in that, The process combines the identified pet species, individual identity, and retrieved individual data to generate a personalized feeding plan that includes feeding time, single feeding amount, and suitable food type, including: Based on the identified pet species, the basic feeding rules for the corresponding species are determined. The feeding frequency, single feeding amount and suitable food type in the basic feeding rules are adjusted in combination with the pet's age, weight and health-related data. The feeding time nodes are optimized by combining the pet's historical eating data, and a unique personalized feeding plan is generated for the pet.

8. The method according to claim 1, characterized in that, The process of controlling the feeder to complete feeding according to the generated personalized feeding plan includes: Based on the generated personalized feeding plan, when the preset feeding time node arrives, the feeding actuator of the feeder is driven to start the feeding action, and the food output data during the feeding process is collected. When the cumulative output reaches the preset threshold corresponding to the single feeding amount of this plan, the operation of the feeding actuator stops, and the single precise feeding is completed.

9. The method according to claim 1, characterized in that, The method further includes: After a single feeding is completed, the visual acquisition module acquires an image of the remaining food in the feeding area, identifies and calculates the actual amount of food consumed by the pet during this feeding, and updates the pet's individual data by combining the feeding time and the pet's eating behavior data. Based on the updated individual data, the subsequent personalized feeding plan is dynamically optimized and adjusted.

10. The method according to claim 1, characterized in that, The method further includes: When the visual acquisition module detects that multiple pets are present in the preset feeding area of ​​the feeder, the species and individual identity of each pet are identified and confirmed. The personalized feeding plan and feeding permissions for each pet are retrieved. The corresponding feeding action is only executed when the target pet that matches the current feeding plan is in the feeding area.

Images