TW202226047A - Method and system of image analysis for infant care - Google Patents

Abstract

A method of image analysis for infant care is provided. The method includes: training, by a managing device, a two-stage model, wherein the two-stage model includes a front-stage model and a back-stage model, and the managing device transmits the front-stage model and the back-stage model; receiving, by a smart front-end device, the front-stage model and inputting at least one image into the front-end model, de-identifying the at least one image to generate at least one de-identified image and transmitting the at least one de-identified image; and receiving, by a back-end server, the back-stage model and the at least one de-identified image, inputting the at least one de-identified image into the back-stage model to determine whether there is an abnormal event in the at least one de-identified image.

Description

Method and system for baby care image analysis

The present disclosure relates to a method and system for image analysis, and in particular, to a method and system for infant care image analysis.

Infants have the highest risk of sudden death within 3 months after birth, and there are many reasons for sudden death. For example, spitting up milk (incidence as high as 15%) can cause choking and suffocation, threatening the life of the baby. However, it is impossible for caregivers to inspect the condition of the baby at any time. Therefore, the risk of sudden infant death can be effectively reduced by using real-time monitoring images, combined with AI image analysis to detect the occurrence of abnormal events.

There are many types of abnormal events that can occur during infant care. If the AI abnormal event detection is performed by the front-end device, better computing performance is required, which often limits the detection accuracy of the abnormal event.

Another solution is to use the AI image analysis function provided by the back-end platform to reduce the computation required by the front-end device. However, in order to perform abnormal event detection through the back-end platform, real-time images need to be transmitted to the back-end platform, resulting in exposure of infant images and the risk of personal data privacy leakage.

The traditional methods of hiding personal information of images, such as mosaic, blurring and other technologies, all use destructive operations on the original image to achieve the purpose of hiding image personal information. However, as a result of such disruptive operations, the backend platform will not be able to perform image-based AI analysis.

Therefore, there is a need for a method and system for analyzing baby care images to improve the above problems.

The following disclosure is exemplary only and is not intended to be limiting in any way. In addition to the illustrated aspects, embodiments, and features, other aspects, embodiments, and features will be apparent by reference to the drawings and the following detailed description. That is, the following disclosure is provided to introduce the concepts, highlights, benefits, and novel and non-obvious technical advantages described herein. Selected, but not all, embodiments are described in further detail below. Accordingly, the following disclosure is not intended to be an essential feature of the claimed subject matter, nor is it intended to be used in determining the scope of the claimed subject matter.

Therefore, the main purpose of the present disclosure is to provide a method and system for analyzing baby care images to improve the above shortcomings.

The present disclosure provides a method for analyzing baby care images, comprising: training a two-stage model by a manager, wherein the two-stage model includes a front-end model and a back-end model, and the manager transmits the front-end model and the back-end model. The above-mentioned back-end model; the above-mentioned front-end model is received by a smart front end, and at least one image is input into the above-mentioned front-end model, and the at least one image is de-identified to generate at least one de-identified image, and the intelligent front end transmits the at least one image. a de-identified image; and receiving the back-end model and the at least one de-identified image by a backend server, and inputting the at least one de-identified image into the back-end model to determine whether the at least one de-identified image exists abnormal event.

In one embodiment, the above-mentioned two-stage model is composed of an N1 hidden layer, a pooling layer, and an N2 hidden layer in series.

In one embodiment, the above-mentioned front-end model system includes the above-mentioned N1 hidden layer and the above-mentioned pooling layer.

In one embodiment, the above-mentioned front-end model system includes the above-mentioned N1 hidden layer and the above-mentioned pooling layer.

In one embodiment, training the two-stage model by the manager includes the following steps: step (a): generating a set of model parameters; step (b): generating the two-stage model according to the set of model parameters; step (c): Detecting abnormal events in a plurality of training images according to the above-mentioned two-stage model, and identifying faces in the above-mentioned plurality of training images; Step (d): According to the above-mentioned detection results and the above-mentioned identification results, determine the above-mentioned Whether the abnormal event detection performance of the two-stage model is greater than a first threshold, and whether the face detection performance is less than a second threshold; Step (e): when the abnormal event detection performance is not greater than the first threshold, Or when the above-mentioned face detection performance is not less than the above-mentioned second threshold, generate another group of model parameters according to the above-mentioned group of model parameters; and replace the above-mentioned group of model parameters with the above-mentioned another group of model parameters, and repeat the above-mentioned steps (b)～( e), until the abnormal event detection performance is greater than the first threshold, and the face detection performance is less than the second threshold.

In one embodiment, the manager generates the other set of model parameters according to the set of model parameters through a heuristic algorithm, wherein the heuristic algorithm is a genetic algorithm, a particle swarm algorithm or a Simulated Annealing Algorithm.

In one embodiment, the manager uses a mean of Average Precision (mAP) method to estimate the abnormal event detection performance and the face detection performance.

In one embodiment, the above-mentioned set of model parameters and the above-mentioned another set of model parameters at least include: the ratio of the number of hidden layers N1 and N2, the depth of each hidden layer, the size of the convolution core of each hidden layer, the activation function of each hidden layer, and the pooling Layer core size.

In one embodiment, the ratio of the number of hidden layers N1 and N2 is 1:7, 1:3 or 3:5.

In one embodiment, the above two-stage model is a Deep Neural Network (DNN) model.

The present disclosure provides a system for analyzing baby care images, including: a manager for training a two-stage model, wherein the two-stage model includes a front-end model and a back-end model, and transmits the front-end model and the back-end model; a an intelligent front-end, receiving the above-mentioned front-end model, inputting at least one image into the above-mentioned front-end model, de-identifying the at least one image, to generate at least one de-identified image, and transmitting the above-mentioned de-identified image; and a backend server, The back-end model and the de-identified image are received, and the at least one de-identified image is input into the back-end model to determine whether there is an abnormal event in the at least one de-identified image.

Aspects of the present disclosure will be described more fully hereinafter with reference to the accompanying drawings. This disclosure may, however, be embodied in many different forms and should not be construed as limited to any specific structure or function presented throughout this disclosure. Rather, these aspects are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art. Based on the teachings herein, one skilled in the art should appreciate that the scope of the present disclosure is intended to encompass any aspect disclosed herein, whether implemented alone or in conjunction with any other aspect of the present disclosure. For example, it may be implemented using any number of means or performing methods set forth herein. Additionally, in addition to the aspects of the disclosure set forth herein, the scope of the disclosure is intended to cover apparatus or methods implemented using other structures, functions, or structures and functions. It should be understood that any aspect disclosed herein may be embodied by one or more elements of the claimed scope.

The word "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any aspect of the present disclosure or designs described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other aspects of the present disclosure or designs. Furthermore, like numerals refer to like elements throughout the several figures, and unless otherwise specified in the description, the articles "a" and "above" include plural references.

It will be understood that when an element is referred to as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements may be present. In contrast, when an element is referred to as being "directly connected" or "directly coupled" to another element, there are no intervening elements present. Other words used to describe the relationship between elements should be interpreted in a similar fashion (eg, "between" versus "directly between," "adjacent" versus "directly adjacent," etc.).

FIG. 1 is a schematic diagram of a baby care image analysis system 100 according to an embodiment of the present invention. The baby care image analysis system 100 includes at least a smart front end 110 , a manager 120 and a backend server 130 .

The smart front-end 110 , the manager 120 , and the back-end server 130 are independent devices, which can be located in different locations, are physically separated, and can be connected to each other through a network. The manager 120 is mainly used to integrate the smart front-end 110 and the back-end server 130 to provide AI image analysis services. The manager 120 includes a neural network training system for training and generating a two-stage model, wherein the two-stage model includes a front-stage model and a back-stage model. Then, the manager 120 can transmit the front-end model to the smart front-end 110 and the back-end model to the backend server 130 .

The smart front end 110 is disposed on a baby care basin 160 , and may include a camera device 112 , a data processor 116 , and a bracket 118 . The photographing device 112 is disposed on the baby care basin 160 through the bracket 118 . The data processor 116 receives at least one image of the baby 114 captured by the photographing device 112, and uses the preceding model received by the smart front end 110 to input the at least one image into the preceding model to de-identify the at least one image to generate at least one image. A de-identified image 150. Then, the smart front end 110 transmits the at least one de-identified image 150 to the backend server 130 .

The backend server 130 can receive at least one de-identified image 150 sent by the smart front end 110 and a backend model sent by the manager 120 . The backend server 130 inputs the at least one de-identified image to the back-end model, so as to determine whether there is an abnormal event in the at least one de-identified image 150 .

Types of smart front-end 110, a manager 120, and a back-end server 130 range from small handheld devices (eg, mobile phones/portable computers) to mainframe systems (eg, mainframe computers). Examples of portable computers include personal digital assistants (PDAs), notebook computers, and the like. The network may include, but is not limited to, one or more Local Area Networks (LANs) and/or Wide Area Networks (WANs).

It should be understood that the smart front end 110 , a manager 120 and a backend server 130 shown in FIG. 1 are examples of the architecture of a system 100 for analyzing baby care images. Each of the elements shown in FIG. 1 may be implemented by any type of computing device, such as computing device 600 described with reference to FIG. 6 , as shown in FIG. 6 .

FIG. 2 is a flowchart illustrating a method 200 of infant care image analysis according to an embodiment of the present disclosure. This method can be implemented in the infant care image analysis system 100 as shown in FIG. 1 .

In step S205, the manager 120 trains to generate a two-stage model, wherein the two-stage model includes a front-end model and a back-end model, and transmits the front-end model to the smart front-end 110 and the back-end model to the backend Server 130. In one embodiment, the above-mentioned two-stage model is a Deep Neural Network (DNN) model, and is composed of N1 hidden layers, a pooling layer, and N2 hidden layers in series.

Next, in step S210, a smart front-end 110 receives at least one image of the baby 114 and the preceding model from the camera 112, inputs the at least one image into the preceding model, and de-identifies the at least one image to obtain At least one de-identified image 150 is generated, and the at least one de-identified image is sent to the background server 130 , wherein the front-end model includes the N1 hidden layer and the pooling layer.

In step S215, a background server 130 receives the de-identified image 150 and the back-end model, and inputs the at least one de-identified image 150 into the back-end model to determine whether there is an abnormal event in the at least one de-identified image 150, The above-mentioned back-end model system includes the above-mentioned N2 hidden layer.

In one embodiment, when the background server 130 determines that there is an abnormal event in at least one de-identified image 150 , the background server 130 may send a warning message to the smart front end 110 to notify the user who operates the smart front end 110 , wherein the smart front end 110 Relevant user interface (eg: Light Emitting Diode (LED), Liquid Crystal Display (LCD), Microphone, Buzzer, Bluetooth Streaming) can be used to remind the user.

FIG. 3 is a structural diagram of a two-stage model 300 according to an embodiment of the present disclosure.

As shown in the figure, the two-stage model 300 is composed of an N1 hidden layer 310 , a pooling layer 320 , and an N2 hidden layer 330 connected in sequence. A hidden layer consists of a convolution (Convolution) layer and an activation (Rectified Linear, Re-Lu) layer. The pooling layer may be a max pooling layer. The convolution kernel size of the hidden layer is k×k, and the kernel size of the pooling layer is p×p.

The front-end model system includes an N1 hidden layer 310 and a pooling layer 320 , and the back-end model system includes an N2 hidden layer 330 . An original image is first input to the N1 hidden layer 310 and the pooling layer 320 . A de-identified image 150 is output through the pooling layer 320 , that is, the image output through the pooling layer 320 will no longer be recognizable. The hidden layer 330 of the N2 layer determines whether there is an abnormal event on the de-identified image. In one embodiment, the abnormal events occurring in the infant may include abnormal events such as eye opening, milk spillage, cyanosis, jaundice, and the like.

The following will describe in detail how the manager 120 trains the above-mentioned two-stage model in step S205 in FIG. 2 . Note that, as used herein, the term "training" is used to identify objects for training a two-stage model. Thus, training images refer to the images used to train the two-stage model.

FIG. 4 shows a schematic diagram 400 of training the above two-stage model by the manager 120 according to an embodiment of the present disclosure, wherein the manager 120 may at least include a model parameter selector, a model trainer, an abnormal event detector, and A face detector.

In block 405, the manager 120 can be initialized first, and then the model parameter selector 410 generates a set of model parameters, wherein the set of model parameters is randomly generated, and the set of model parameters at least includes: the number of hidden layers N1 and The ratio of N2, the depth of each hidden layer, the size of the convolution core of each hidden layer, the activation function of each hidden layer, and the core size of the pooling layer.

Next, the above-mentioned set of model parameters are input into the model trainer 415 to generate a two-stage model 420, the two-stage model 420 includes a front-end model and a back-end model, wherein the front-end model is composed of N1 hidden layers and the above-mentioned pooling layer, The back-end model is composed of N2 hidden layers. Next, a plurality of training images are input into the two-stage model 420, a plurality of de-identified training images are generated by the previous model, and the plurality of de-identified training images are output to the face detector 435, and the face detector The detector 435 identifies the faces to which the plurality of de-identified training images correspond respectively. The latter-stage model will receive the plurality of de-identification training images, and output the plurality of abnormal de-identification training images to the abnormal event detector 425 , and the abnormal event detector 425 will identify why the plurality of abnormal de-identification training images are described above. The abnormal event is then input to the abnormal event detector performance estimator 430 for evaluation.

For the abnormal events and faces corresponding to each training image, the manager 120 has, for example, a comparison table for the abnormal event detector performance estimator 430 to estimate the abnormal event detection performance, that is, the correct rate of abnormal event recognition, And for the face detection performance estimator 440 to estimate the face detection performance, that is, the correct rate of face recognition for the de-identified image. Finally, in block 445, the manager 120 determines whether the performance estimated by the abnormal event detector performance estimator 430 and the face detection performance estimator 440 meets the target condition (the abnormal event detection performance is greater than a first threshold, and whether the above-mentioned face detection performance is less than a second threshold).

When the abnormal event detection performance and the face detection performance meet the target conditions, in block 450, the manager ends the training. When the abnormal event detection performance and the face detection performance meet the target conditions, in block 450, the manager ends the training.

When the abnormal event detection performance and the face detection performance do not meet the target conditions, the manager will output the above results to block 410, so that the model parameter selector can generate a new set of models according to the above set of model parameters parameter.

FIG. 5 shows a flowchart 500 of training the above-mentioned two-stage model by the manager 120 according to an embodiment of the present disclosure, which further illustrates the details of the process in FIG. 4 .

In step S505, the manager 120 generates a set of model parameters, wherein the set of model parameters is generated by random initialization, and the set of model parameters at least includes: the ratio of the number of hidden layers N1 and N2, the depth of each hidden layer, the depth of each hidden layer The size of the convolution kernel, the activation function of each hidden layer, and the kernel size of the pooling layer.

In step S510, the manager 120 generates a two-stage model according to the above-mentioned set of model parameters. Next, in step S515, the manager 120 detects abnormal events in the plurality of training images according to the above-mentioned two-stage model, and recognizes the human faces in the above-mentioned plurality of training images.

Next, in step S520, the manager 120 determines whether the abnormal event detection performance of the two-stage model is greater than a first threshold and whether the face detection performance is less than a second threshold. In more detail, the abnormal event detection performance should be as large as possible, indicating that the performance of the back-end model in detecting abnormal events is better. The face detection performance should be as small as possible, which means that the back-end model cannot identify the face image from the de-identified image generated by the front-end model. In one embodiment, the manager uses a mean of Average Precision (mAP) method to estimate the abnormal event detection performance and the face detection performance. In another embodiment, the first threshold and the second threshold are pre-specified manually.

When the abnormal event detection performance is not greater than the first threshold, or the face detection performance is not less than the second threshold (“No” in step S520 ), in step S525 , the manager 120 according to the above-mentioned group model The parameters generate another set of model parameters, wherein the above-mentioned another set of model parameters at least includes: the ratio of the number of hidden layers N1 and N2, the depth of each hidden layer, the size of the convolution core of each hidden layer, the activation function of each hidden layer, and the core size of the pooling layer . In one embodiment, the manager 120 generates the other set of model parameters according to the set of model parameters through a heuristic algorithm, wherein the heuristic algorithm is a genetic algorithm, a particle swarm optimization or A simulated annealing algorithm and other algorithms.

After the manager 120 generates another set of model parameters according to the set of model parameters, the set of model parameters is replaced by the other set of model parameters, and the above steps S510 to S525 are repeatedly executed until the abnormal event detection performance is greater than the first set of model parameters. threshold, and the above-mentioned face detection performance is less than the above-mentioned second threshold. The two-stage model trained after steps S510 to S525 will be divided into a front-stage model and a back-stage model. The front-end model and the back-end model are further sent to the smart front-end 110 and the back-end server 130 for execution, respectively.

。 In one embodiment, in order to be suitable for the combination optimization method, the model parameters must be designed as a finite combination, for example, the ratio of the number of hidden layers N1 and N2 is (1:7, 1:3, 3:5) , the size of the convolution kernel (K) of each hidden layer is (3×3, 4×4, 5×5, 6×6), the depth of the hidden layer (C) is (4, 16, 32, 64, 128), The core size of the pooling layer is (2×2, 3×3, 4×4), and the model parameter model can be expressed as follows:

.

As described above, the baby care image analysis method and system of the present disclosure divides the image analysis into multiple segments, and can support both local and cloud-based architectures, generate non-destructive and private images through the front-end model, and then use the back-end model to judge Whether there is an abnormal event in the private image. In other words, the method and system for analyzing baby care images of the present disclosure utilizes a deep neural network (DNN) to change the characteristics of the original image, and the original image is processed by the front-end model as a means of hiding personal information. In addition, the baby care image analysis method and system disclosed in the present disclosure further proposes a DNN model segmentation method, which can quantify the extent to which the front-end model hides personal information and generate segmentation point recommendations, which can effectively reduce the computational workload of the front-end computing device, and can Avoid the risk of personal data privacy leakage when the back-end platform is operating.

For the described embodiments of the invention, an exemplary operating environment in which embodiments of the invention may be implemented is described below. Referring specifically to FIG. 6, an exemplary operating environment, which may generally be considered a computing device 600, is shown for implementing embodiments of the present invention. Computing device 600 is merely one example of a suitable computing environment and is not intended to imply any limitation on the scope of use or functionality of the present invention. Neither should computing device 600 be interpreted as having any dependency or requirement relating to any one or combination of the elements shown.

The invention may be implemented in computer code or machine-useable instructions, the instructions may be computer-executable instructions of a program module, the program module of which is executed by a computer or other machine, such as a personal digital assistant or other portable device . Generally speaking, a program module includes routines, programs, objects, components, data structures, etc. A program module refers to a program code that performs a specific task or implements a specific abstract data type. The present invention can be implemented in a variety of system configurations, including portable devices, consumer electronics, general purpose computers, more specialized computing devices, and the like. The present invention can also be implemented in a distributed computing environment, processing devices connected by a communication network.

Refer to Figure 6. Computing device 600 includes a bus 610, memory 612, one or more processors 614, one or more display elements 616, input/output (I/O) ports 618, input/output (I/O) ports 618, which are directly or indirectly coupled to Output (I/O) elements 620 and illustrative power supply 622 . Bus 610 represents an element (eg, an address bus, a data bus, or a combination thereof) that may be one or more bus bars. Although each block of FIG. 6 is shown as a line for brevity, in fact, the demarcation of each element is not specific, for example, the presentation element of the display device may be regarded as an I/O element; the processor may have a memory .

Computing device 600 generally includes various computer-readable media. Computer-readable media can be any available media that can be accessed by computing device 600, including both volatile and non-volatile media, removable and non-removable media. By way of example and not limitation, computer-readable media may include computer storage media and communication media. Computer-readable media includes both volatile and non-volatile media implemented in any method or technology for storage of information such as computer-readable instructions, data structures, program modules or other data, Removable and non-removable media. Computer storage media include but are not limited to (Random Access Memory, RAM), Read-Only Memory (ROM), Electronically-Erasable Programmable Read-Only Memory (EEPROM, EEPROM) ), flash memory or other memory technology, CD-ROM, Digital Versatile Disc (DVD) or other optical disk storage device, magnetic disk, magnetic disk, platter storage device or other magnetic storage device, or Any other medium that can be used to store the required information and that can be accessed by the computing device 600 . The computer storage medium itself does not contain the signal.

Communication media typically include computer-readable instructions, data structures, program modules, or other data in the form of modular data signals such as carrier waves or other transport mechanisms, and include any information delivery media. The term "modular data signal" refers to a signal that has one or more sets of features or is modified in a manner that encodes information in the signal. By way of example and not limitation, communication media include wired media such as a wired network or direct wired connection and wireless media such as audio, radio frequency, infrared, and other wireless media. Combinations of the above are included within the scope of computer-readable media.

Memory 612 includes computer storage media in the form of volatile and non-volatile memory. The memory can be removable, non-removable, or a combination of the two. Exemplary hardware devices include solid state memory, hard disk drives, optical disk drives, and the like. Computing device 600 includes one or more processors that read data from entities such as memory 612 or I/O element 620 . Display element 616 displays data indications to a user or other device. Exemplary display elements include display devices, speakers, printing elements, vibrating elements, and the like.

I/O ports 618 allow computing device 600 to logically connect to other devices including I/O elements 620, some of which are built-in devices. Exemplary elements include microphones, joysticks, game consoles, satellite dishes, scanners, printers, wireless devices, and the like. I/O element 620 may provide a natural user interface for processing user-generated gestures, sounds, or other physiological inputs. In some instances, these inputs can be passed to an appropriate network element for further processing. Computing device 600 may be equipped with depth cameras, such as stereo camera systems, infrared camera systems, RGB camera systems, and combinations of these systems, to detect and identify objects. In addition, the computing device 600 may be equipped with sensors (eg, radar, lidar) to periodically sense the surrounding environment within a sensing range, and generate sensor information representing the relationship between itself and the surrounding environment. Furthermore, computing device 600 may be equipped with an accelerometer or gyroscope that detects motion. The output of the accelerometer or gyroscope may be provided to computing device 600 for display.

In addition, the processor 614 in the computing device 600 can also execute programs and instructions in the memory 612 to perform the actions and steps described in the above embodiments, or other descriptions in the specification.

Any specific order or layered steps of the procedures disclosed herein are by way of example only. Based on design preferences, it is understood that any specific order or layered steps in the procedure may be rearranged within the scope disclosed in this document. The accompanying method claims present elements of the various steps in a sample order, and are therefore not to be limited by the specific order or hierarchy presented.

The use of ordinal numbers such as "first", "second" and "third" used to modify elements in the scope of the patent application do not imply any priority, priority order, order between elements, or method execution. The order of the steps is only used as an indicator to distinguish different elements with the same name (with different ordinal numbers).

Although the present disclosure has been disclosed above with examples, it is not intended to limit the present case. Anyone who is familiar with the art can make some changes and modifications without departing from the spirit and scope of the present disclosure. Therefore, the protection scope of this case should be The scope of the patent application attached shall prevail.

100: Baby Care Image Analysis System 110: Smart Front End 112: Photographic installations 114: Baby 116: Data Processor 118: Bracket 120:Manager 130: Backend server 150: De-identified images 160:Baby Nursing Basin 200: Method S205, S210, S215: Steps 300: Two-Stage Model 310: Hidden Layer 320: Pooling layer 330: Hidden Layer 400: Schematic 405,445,450: Blocks 410: Model parameter selector 415: Model Trainer 420: Two-Stage Model 425: Abnormal Event Detector 430: Anomaly Event Detection Effectiveness Estimator 435: Face Detector 440: Face Detection Performance Estimator 500: Method S505, S510, S515, S520, S525: Steps 600: Computing Devices 610: Busbar 612: Memory 614: Processor 616: Display Components 618: I/O port 620: I/O Components 622: Power supply

FIG. 1 is a schematic diagram showing the environment of a system for analyzing baby care images according to an embodiment of the present invention. FIG. 2 is a flow chart showing a method for analyzing baby care images according to an embodiment of the present disclosure. FIG. 3 is a structural diagram of a two-stage model according to an embodiment of the present disclosure. FIG. 4 is a schematic diagram of training the above-mentioned two-stage model by a manager according to an embodiment of the present disclosure. FIG. 5 shows a flow chart of a manager training the above-mentioned two-stage model according to an embodiment of the present disclosure. Figure 6 shows an exemplary operating environment for implementing embodiments of the present invention.

200: Method

S205, S210, S215: Steps

Claims (20)

A method of infant care image analysis comprising: A two-stage model is trained by a manager, wherein the two-stage model includes a front-end model and a back-end model, and the front-end model and the back-end model are transmitted by the manager; Receiving the above-mentioned front-end model by a smart front-end, inputting at least one image into the above-mentioned front-end model, de-identifying the at least one image to generate at least one de-identified image, and transmitting the at least one de-identified image by the intelligent front-end images; and A background server receives the back-end model and the at least one de-identified image, and inputs the at least one de-identified image into the back-end model to determine whether the at least one de-identified image has an abnormal event. The method for analyzing baby care images according to claim 1, wherein the two-stage model is composed of an N1 hidden layer, a pooling layer, and an N2 hidden layer in series. The method for analyzing baby care images according to claim 1, wherein the front-end model comprises the N1 hidden layer and the pooling layer. The method for analyzing baby care images according to claim 1, wherein the back-end model includes the N2 hidden layer. The method for analyzing baby care images according to claim 1, wherein training the two-stage model by the manager includes the following steps: Step (a): generating a set of model parameters; Step (b): generate the above-mentioned two-stage model according to the above-mentioned group model parameters; Step (c): Detecting abnormal events in a plurality of training images according to the above-mentioned two-stage model, and identifying faces in the above-mentioned plurality of training images; Step (d): according to the above-mentioned detection results and the above-mentioned identification results, determine whether an abnormal event detection performance of the above-mentioned two-stage model is greater than a first threshold value, and whether the face detection performance is less than a second threshold value; Step (e): when the above-mentioned abnormal event detection performance is not greater than the above-mentioned first threshold, or when the above-mentioned human face detection performance is not less than the above-mentioned second threshold, generate another group of model parameters according to the above-mentioned group of model parameters; and Replace the above-mentioned set of model parameters with the above-mentioned another set of model parameters, and repeat the above-mentioned steps (b) to (e) until the above-mentioned abnormal event detection performance is greater than the above-mentioned first threshold value, and the above-mentioned face detection performance is less than the above-mentioned second threshold value. until. The method for analyzing baby care images according to claim 1, wherein the manager generates the other set of model parameters according to the set of model parameters by a heuristic algorithm, wherein the heuristic algorithm is a genetic algorithm method, a particle swarm optimization algorithm, or a simulated annealing algorithm. The method for analyzing baby care images as claimed in claim 1, wherein the manager uses a mean of Average Precision (mAP) method to estimate the abnormal event detection performance and the face detection performance. The method for analyzing baby care images according to claim 1, wherein the above-mentioned set of model parameters and the above-mentioned another set of model parameters at least include: the ratio of the number of hidden layers N1 and N2, the depth of each hidden layer, and the size of the convolution kernel of each hidden layer , the activation function of each hidden layer, and the core size of the pooling layer. The method for analyzing baby care images according to claim 1, wherein the ratio of the number of hidden layers N1 and N2 is 1:7, 1:3 or 3:5. The method for analyzing baby care images according to claim 1, wherein the two-segment model is a Deep Neural Network (DNN) model. A system for baby care image analysis, comprising: a manager for training a two-stage model, wherein the two-stage model includes a front-end model and a back-end model, and transmits the front-end model and the back-end model; an intelligent front end that receives the preceding model, inputs at least one image into the preceding model, de-identifies the at least one image to generate at least one de-identified image, and transmits the de-identified image; and A backend server receives the back-end model and the de-identified image, and inputs the at least one de-identified image into the back-end model to determine whether there is an abnormal event in the at least one de-identified image. The system for analyzing baby care images according to claim 11, wherein the two-stage model is composed of an N1 hidden layer, a pooling layer and an N2 hidden layer in series. The system for analyzing baby care images according to claim 12, wherein the front-end model includes the N1 hidden layer and the pooling layer. The system for analyzing baby care images according to claim 12, wherein the back-end model includes the N2 hidden layer. The system for analyzing baby care images according to claim 11, wherein the training of the two-stage model by the manager includes the following steps: Step (a): generating a set of model parameters; Step (b): generate the above-mentioned two-stage model according to the above-mentioned group model parameters; Step (c): train the above two-stage model to detect abnormal events in the plurality of training images, and identify the faces in the above plurality of training images; Step (d): according to the above-mentioned detection results and the above-mentioned identification results, determine whether an abnormal event detection performance of the above-mentioned two-stage model is greater than a first threshold value, and whether the face detection performance is less than a second threshold value; Step (e): when the above-mentioned abnormal event detection performance is not greater than the above-mentioned first threshold, or when the above-mentioned human face detection performance is not less than the above-mentioned second threshold, generate another group of model parameters according to the above-mentioned group of model parameters; and Replace the above-mentioned set of model parameters with the above-mentioned another set of model parameters, and repeat the above-mentioned steps (b) to (e) until the above-mentioned abnormal event detection performance is greater than the above-mentioned first threshold value, and the above-mentioned face detection performance is less than the above-mentioned second threshold value. until. The system for analyzing baby care images as claimed in claim 15, wherein the manager generates the other set of model parameters according to the set of model parameters through a heuristic algorithm, wherein the heuristic algorithm is a genetic algorithm method, a particle swarm optimization algorithm, or a simulated annealing algorithm. The system for analyzing baby care images as claimed in claim 15, wherein the manager uses a mean of Average Precision (mAP) method to estimate the abnormal event detection performance and the face detection performance. The system for analyzing baby care images according to claim 15, wherein the above-mentioned set of model parameters and the above-mentioned another set of model parameters at least include: the ratio of the number of hidden layers N1 and N2, the depth of each hidden layer, the convolution core of each hidden layer size, the activation function of each hidden layer, and the core size of the pooling layer. The system for analyzing baby care images according to claim 15, wherein the ratio of the hidden layers N1 and N2 is 1:7, 1:3 or 3:5. The system for analyzing baby care images according to claim 11, wherein the two-stage model is a Deep Neural Network (DNN) model.

Images