WO2022237061A1

WO2022237061A1 - Embedded object cognitive system based on image processing

Info

Publication number: WO2022237061A1
Application number: PCT/CN2021/122781
Authority: WO
Inventors: 王宜怀; 刘纯平; 王进; 施连敏; 胡展鹏; 常诚
Original assignee: 苏州大学
Priority date: 2021-05-10
Filing date: 2021-10-09
Publication date: 2022-11-17
Also published as: CN113158968A

Abstract

An embedded object cognitive system (100) based on image processing. The embedded object cognitive system (100) based on image processing comprises an image acquisition module (10), a model training module (20), a model terminal deployment module (30), and a terminal reasoning module (40). The system can collect images in real time and obtain the recognition result after reasoning by means of a lightweight convolutional neural network model in the model training module (20). The system can greatly reduce the demand for hardware resources while ensuring the accuracy of object recognition and the reasoning speed, and realizes the recognition and classification of different types of objects.

Description

An Embedded Object Recognition System Based on Image Processing

technical field

The invention relates to the technical field of embedded artificial intelligence, in particular to an embedded object recognition system based on image processing.

Background technique

Embedded Artificial Intelligence (EAI) is the product of deep integration of embedded computer technology, artificial intelligence technology and the actual needs of each application scenario. In addition to the technical advantages of artificial intelligence, embedded artificial intelligence also has the excellent real-time performance, applicability, robustness and stability of embedded technology.

The traditional embedded intelligent software and hardware platform centers on the cloud server, and transmits the raw data collected by the terminal to the cloud. Data storage and analysis are all completed in the cloud, and the embedded terminal only realizes data collection and output results. The corresponding operation completes a cycle of data in this way. This kind of intelligent embedded software and hardware platform with cloud computing as the core has problems such as high overhead, poor real-time performance, and data privacy, and cannot meet most practical application needs. With the development of technology, the processing capability of the terminal is becoming more and more powerful, and emerging technologies such as edge computing and fog computing aiming to overcome the shortcomings of the embedded intelligent software and hardware platform at the core of the cloud have been proposed. However, this type of method only divides the propagation process of the network model into two parts: the terminal and the cloud. The embedded terminal does not have complete cognitive capabilities, and the complete reasoning process still requires cloud computing.

Therefore, in view of the above technical problems, it is necessary to provide an embedded object recognition system based on image processing. The system collects images in real time, and obtains recognition results after reasoning through a lightweight convolutional neural network model in the model training module; the system can greatly reduce the demand for hardware resources while ensuring the accuracy of object recognition and reasoning speed , to realize the recognition and classification of different types of objects.

technical solution

In view of this, the purpose of the embodiments of the present invention is to provide an embedded object recognition system based on image processing, which collects images in real time and obtains recognition results after reasoning through a lightweight convolutional neural network model in the model training module ; The system can greatly reduce the demand for hardware resources while ensuring the accuracy of object recognition and reasoning speed, and realize the recognition and classification of different types of objects.

In order to achieve the above object, the technical solution provided by the embodiment of the present invention is as follows: an embedded object recognition system based on image processing includes: an image acquisition module, used to collect image features of training objects; a model training module, connected with the image The acquisition module is connected, and the image features obtained by the image acquisition module are used as training materials, and a preset algorithm is used to generate a cognitive model parameter component that can be directly compiled and used under the embedded engineering framework; the model terminal deployment module is used for Deploy the cognitive model parameter components obtained by the model training module on the embedded terminal; the terminal reasoning module uses the cognitive model provided by the model terminal deployment module according to the image of the target object collected by the image collection module The parameter component performs target object recognition.

As a further improvement of the present invention, the model training module includes two modes: a PC model training mode and an embedded terminal real-time reasoning training mode.

As a further improvement of the present invention, the PC model training mode includes the steps: the embedded terminal acquires image features and transmits the image features to the PC; the PC builds a data set according to the image features and follows the presets in the model training module The algorithm generates a cognitive model parameter component that can be directly compiled and used under the embedded engineering framework; the cognitive model parameter component is deployed to the embedded terminal by burning.

As a further improvement of the present invention, the real-time inference training mode of the embedded terminal includes the steps: the embedded terminal acquires image features; the terminal inference module performs image processing according to the image features and generates an image that can be used according to the preset algorithm in the model training module. A cognitive model parameter component directly compiled and used under the embedded engineering framework; the cognitive model parameter component is stored in an embedded terminal.

As a further improvement of the present invention, the hardware configuration in the model terminal deployment module is different according to the nature of the parameters in the embedded object recognition system.

As a further improvement of the present invention, constant parameters are stored in FLASH memory, and variable parameters are stored in RAM memory.

As a further improvement of the present invention, the constant parameters include filter erosion, bias BIAS parameters, and propagation structure functions; the variable parameters include image features, input variables and output variables.

As a further improvement of the present invention, the embedded object recognition system replaces the reading and writing of the dynamic array in the RAM with the erasing and reading and writing of the continuous address space of the preset FLASH designated sector.

As a further improvement of the present invention, the parameter format used in the preset algorithm in the model training module is a multidimensional array form in C language.

As a further improvement of the present invention, an optimized camera driving algorithm is used in the image acquisition module to drive the camera to acquire image features of the training object.

As a further improvement of the present invention, the image acquisition module includes an image processing unit, and the image processing unit uses a threshold filtering method to process the original image of the training object obtained by the image acquisition module to obtain the image features of the training object .

As a further improvement of the present invention, the model training module uses a fusion rolling convolution algorithm to generate a cognitive model parameter component that can be directly compiled and used under the embedded engineering framework.

Beneficial effect

The present invention has the following advantages:

The purpose of the embodiments of the present invention is to provide an embedded object recognition system based on image processing, which collects images in real time and obtains recognition results after reasoning through a lightweight convolutional neural network model in the model training module. Furthermore, the model training module in the system adopts the fusion rolling convolution algorithm to effectively optimize the embedded system's demand for the size and space of the image area. Furthermore, the system uses an optimized camera driving algorithm to drive the camera during the image acquisition process, which effectively improves the speed of image reading and display. The system can greatly reduce the demand for hardware resources while ensuring the accuracy of object recognition and reasoning speed, and realize the recognition and classification of different types of objects.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments described in the present invention. Those skilled in the art can also obtain other drawings based on these drawings without creative work.

FIG. 1 is a schematic diagram of modules of an embedded object recognition system provided by an embodiment of the present invention;

Fig. 2 is a schematic diagram of the data flow model of the embedded object recognition system of the embodiment shown in Fig. 1;

Fig. 3 (a) is the schematic flow chart of PC model training mode in the embodiment shown in Fig. 1;

Fig. 3 (b) is the schematic flow chart of embedded terminal real-time inference training mode in the embodiment shown in Fig. 1;

Fig. 4 is a frame diagram of hardware configuration allocation in the model terminal deployment module in the embodiment shown in Fig. 1;

Fig. 5 is the flowchart of optimizing camera driving algorithm in the embodiment of the present invention;

6(a), 6(b), and 6(c) are schematic diagrams of the fusion rolling convolution algorithm in the embodiment of the present invention.

Explanation of reference signs:

100. Embedded object recognition system 10. Image acquisition module 20. Model training module 30. Model terminal deployment module 40. Terminal reasoning module.

Embodiments of the present invention

In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described The embodiments are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

As shown in FIG. 1 , a block diagram of an embedded object recognition system based on image processing provided by an embodiment of the present invention. In this embodiment, the embedded object recognition system 100 based on image processing includes an image acquisition module 10 , a model training module 20 , a model terminal deployment module 30 and a terminal reasoning module 40 . The image acquisition module 10 is used to acquire image features of the training object. The model training module 20 is connected with the image acquisition module 10, uses the image features obtained by the image acquisition module 10 as the training material, and adopts a preset algorithm to generate cognitive model parameter components that can be directly compiled and used under the embedded engineering framework. The model terminal deployment module 30 is used to deploy the cognitive model parameter components obtained by the model training module 20 on the embedded terminal. The terminal inference module 40 recognizes the target object according to the image of the target object collected by the image collection module 10 and uses the cognitive model parameter components provided by the model terminal deployment module 30 .

Continuing to refer to FIG. 2 , the image acquisition module 10 first collects the feature data of the corresponding object as the material for training the cognitive model; after collecting a sufficient amount of image data, the model training module 20 trains the model, and finally generates an image that can be embedded in a general purpose through a related algorithm. Cognitive model parameter components directly compiled and used under the engineering framework; the model terminal deployment module 30 deploys the cognitive model parameter components obtained by the model training module 20 on the embedded terminal, that is, after recompiling and burning, the new cognitive model The cognitive model is deployed on the terminal, and at this time, the terminal can recognize the target object through the terminal reasoning module 40 to obtain a cognitive result.

In the overall application system, the model training module 20 can be divided into two modes: PC model training mode and embedded terminal real-time reasoning training mode according to the direction of data transmission and the different functions performed by the system.

Among them, the flow of the PC model training mode is shown in Figure 3(a). The PC model training mode includes steps: the embedded terminal obtains image features and transmits the image features to the PC end, that is, the image acquisition module 10 collects image feature data of the corresponding object and transmits the image features to the PC end; Create a data set and generate a cognitive model parameter component that can be directly compiled and used under the embedded engineering framework according to the preset algorithm in the model training module; deploy the cognitive model parameter component to the embedded terminal by burning .

Among them, the flow of the real-time inference training mode of the embedded terminal is shown in Fig. 3(b). The embedded terminal real-time inference training mode includes steps: the embedded terminal obtains image features, that is, the image acquisition module 10 collects image feature data of corresponding objects; the terminal inference module performs image processing according to the image features and follows the preset in the model training module The algorithm generates cognitive model parameter components that can be directly compiled and used under the embedded engineering framework; the cognitive model parameter components are stored in the embedded terminal.

The physical resource memory of the MCU in the embedded system is divided into two types: volatile memory and non-volatile memory. Random Access Memory (Random Access Memory, RAM) and flash memory (Flash EEPROM Memory, FLASH) are the representative devices of the two respectively. RAM provides temporary data such as local variables required by the processor during operation, and FLASH stores read-only data for program operation and the program itself. Under normal circumstances, the space size of RAM is much smaller than that of FLASH. Taking the STM32L431RC chip as an example, the size of RAM is 64KB, while the size of FLASH is 256KB.

The size of the main control chip RAM determines the performance of the system to process data, and in the reasoning process of the network model, the temporary data generated by each sub-network during the reasoning process will not affect the follow-up, and the feed-forward neural network between each layer The layer-by-layer propagation of the network model is realized only by outputting the feature matrix, so the total data volume of all parameters used by the single-layer network should not be greater than the RAM space size of the main control chip. The resource consumption of the network model in the embedded terminal is also divided into the deployed parameter model itself and the additional data consumption generated by the runtime model propagation. In the forward propagation process of the network model from input to output, the data resources involved are the input image, network model parameters, temporary space generated during the operation process, input and output during each layer transfer process, and the output of the final model. There are five data categories.

In this embodiment, the image processing-based embedded object recognition system 100 designs a reasonable model resource configuration framework according to the spatial resource characteristics of the embedded terminal chip. The hardware configuration in the model terminal deployment module 30 is different according to the nature of the parameters in the embedded object recognition system. Specifically, the constant parameters are stored in the FLASH memory, and the variable parameters are stored in the RAM memory. Adapt the model parameters of different characteristics to different physical storage resources in the embedded terminal, rationalize the resource allocation method for terminal operation, and reduce unnecessary resource consumption.

As shown in FIG. 4 , in this embodiment, the constant parameters include filter erosion, bias BIAS parameters, and propagation structure functions; the variable parameters include image features, input variables, and output variables. For network model parameters that do not change during propagation, such as convolution kernel parameters, bias, etc., since network model parameters usually occupy the largest storage space and cannot be updated and iterated during network model reasoning, the system uses a constant array It is stored in FLASH in the form of FLASH. Compared with RAM, FLASH has a larger resource space and is more suitable for storing such data. Due to its fixed structure, the input and output of each layer of network transmission process exists as a fixed-size multi-dimensional array in the embedded terminal, but the value of each array member will change during the transmission process, so the system uses a dynamic array The method is stored in RAM, which is more conducive to the computing speed of the system. Since the input image and the output array are jointly processed with other software layer components, the system stores them in RAM as global arrays.

In embedded systems, the size of RAM is relatively small, but RAM is responsible for the main data calculations. At the same time, in the process of network model reasoning, the resources occupied by each layer are different, and the space resources occupied by different network structures of each layer are quite different. Taking the VGG16 [54] network as an example, the network parameters and input and output used in the first convolutional layer occupy nearly 100KB of space, while the final fully connected neural network layer only occupies less than 1KB of computing space. How to cut the network layer that takes up the most resources and reduce the space occupied by this layer network by other means is one of the problems that need to be solved urgently.

Preferably, the embedded object recognition system 100 replaces the reading and writing of the dynamic array in the RAM with the erasing and reading and writing of the continuous address space of the preset FLASH designated sector. In this way, the consumption of RAM resources by the embedded object recognition system 100 is reduced, the applicability of the main control chip to the size of the network model is improved, and the accuracy of model reasoning is improved with less time loss. The specific algorithm of erasing and replacing is shown in Table 1:

.

General low-resource embedded terminals cannot directly support neural network algorithm libraries such as Keras due to different compilation frameworks. In this embodiment, the embedded object recognition system 100 first reads the parameters of the H5 format file generated by training on the network model fitting platform, and then generates a C of the model parameters that can be directly compiled in the embedded engineering framework according to the designed algorithm. Language version widget.

Through the data analysis of the H5 file, it is found that the data format in the H5 file is a tree structure, which is divided into two types of data: weight and bias. Among them, the expression form of the specific data of the convolution kernel element a in the h row w column of the nth dimension of the l-layer network is shown in formula 1:

.

The specific expression of the data of the kth bias item b of the l-layer network is shown in Equation 2:

.

Accordingly, the location of the data can be directly located, which provides a theoretical basis for the design of the preset algorithm in the embodiment of the present invention. The embedded object recognition system 100 designs a model reasoning parameter format conversion algorithm to convert the parameters used in model reasoning into a multidimensional array in C language in storage form, and convert the algorithm model into an embedded engineering component. That is, the parameter format used in the preset algorithm in the model training module 20 is a multidimensional array form in C language.

Preferably, the embodiment of the invention first converts the model parameters into a multidimensional array form in C language, then extracts the commonality in the embedded project, adds the attached header file components, and elements such as variable declarations required at the beginning of the file, and converts them into general The form of embedded engineering components can be directly deployed on embedded terminals, providing model parameter data support for terminal reasoning.

Compared with the traditional environmental sensor acquisition process, the image data acquisition process is more complicated, and the amount of tediously processed data is also larger. To address these problems, the embedded object recognition system 100 designs an image acquisition acceleration algorithm and a feature extraction algorithm to ensure that the terminal can quickly acquire high-quality image data. In a preferred embodiment, an optimized camera driving algorithm is used in the image acquisition module 10 to drive the camera to acquire image features of the training object.

Taking the image format QVGA as an example, the size of the image transmitted by the camera to the cache chip is 80×60 pixels, and the pixel information data needs to be read 4800 times, so the method of judging the validity of each pixel and the communication process between the camera and the cache chip The speed increase after zooming in is considerable. As shown in FIG. 5 , the flow chart of optimizing the camera driving algorithm in the embodiment of the present invention. The embedded object recognition system 100 needs to output 2 clock signals in total to obtain a pixel point, that is, 4 GPIO port high/low level signals, and 19,200 clock signals need to be output to read a complete QVGA image. Therefore, the system first performs an acceleration optimization operation for the operation of outputting the clock signal.

The traditional embedded GPIO operation usually calls the packaged GPIO interface function, and the interface function often places parameter verification and judgment operations in the function to ensure the robustness of the program. For example, STMicroelectronics' HAL bottom packaging library, a GPIO pull-up operation requires a series of operations such as port correctness verification and GPIO register verification. These operations avoid irregular parameter input and ensure the reusability and robustness of the function. However, it is not suitable for single GPIO operation of image acquisition, and it also increases the burden of system operation. The embedded object recognition system 100 performs setting operations on the input and output registers corresponding to the communication GPIO port of the camera module, avoiding parameter transmission and judgment, and improving the efficiency and speed of image acquisition.

Crop and compress the pixel-sized image captured by the camera unit in the image capture module 10 . In this way, the space occupied by the image data can be reduced while ensuring that the image feature data is not lost as much as possible. Different from the traditional image processing algorithm for two-dimensional array processing, because the image data acquisition module transmits pixel by pixel, so in the process of image acquisition by the embedded terminal, related operations are performed on the one-dimensional image array. Assume that the image input dimension of the image processing algorithm adopted is H×H pixel size, and the algorithm compresses the image data of QVGA size into H×H format size. In this embodiment, the compression algorithm first cuts the collected image to a size of 60×60, and then judges whether the ordinal number n of the input pixel point is the target pixel point: if so, store it in the corresponding target two-dimensional array position Ax,y , otherwise discard. The specific compression algorithm is shown in Table 2:

.

Further, the embodiment of the present invention also performs LCD display acceleration. The process of displaying single pixel data on the LCD is also the process of the chip and the LCD through the serial peripheral interface (Serial Peripheral Interface). Peripheral Interface, SPI) communication process. Change the process of point-by-point display to first set the LCD display area, and then call the SPI to directly send the pixel data to the LCD, eliminating the need to set the coordinate process for each transmission. In order to maximize the use of MCU resources, the LCD display adopts a point-by-point display method, that is, after receiving a pixel data, it will be displayed directly on the LCD, and only a single pixel resource is occupied and reused to display a pixel on the LCD. a complete image. At the same time, the traditional LCD display pixel point function is to position and display each pixel point, that is, first determine the relative position displayed on the LCD, and then display a corresponding point. But this is inefficient for the image display case where the designated area is repeated many times. Each pixel displayed in the image has a certain positional continuity relationship with the pixels displayed before and after, and it is not necessary to locate each pixel to display a complete image.

Preferably, the image acquisition module 10 further includes an image processing unit (not shown in the figure). The image processing unit uses a threshold filtering method to process the original image of the training object obtained by the image acquisition module to obtain the image features of the training object, thereby effectively filtering out the image background and retaining the image features of the target object itself as much as possible.

In this embodiment, the threshold filtering method specifically includes an edge mean method, a bimodal mean method, and a bimodal valley method. The edge mean method performs the mean value operation on all the edge pixels, and the obtained mean value is used as a threshold to filter the image. The algorithm of bimodal mean method is based on the idea of iterative update. The algorithm first processes the number of occurrences of each gray value in the input image into the form of a histogram, and then performs a bimodal judgment on the histogram, that is, whether there are and only two local maxima appear, and if so, take this The average of the two local maxima is used as the filtering threshold. Otherwise, each data point is smoothed with a span of n, and the number of smoothing is given at the same time. If it exceeds the upper limit N, it is judged that the graph cannot be filtered. After obtaining the gray values of the two maximum frequencies, the bimodal valley method does not take the two intermediate gray values, but takes the lowest valley between the two peaks, that is, the gray value that appears between the two gray values. A gray value with the lowest frequency is used as a threshold to filter the image.

The usual network model processing method is often to obtain the entire image data first, and then input the image data into the network, which is reasonable in the case of high resources. In the case of low resources, if the input image is large, such as a 224×224 pixel image usually used by the network model, a single image occupies a space of 49KB. In this ebb and flow system space, this also affects the number of model parameters, which ultimately reduces the recognition accuracy of the system. How to minimize the space occupied by the input image is the key to improving the space utilization efficiency of the system. To solve this problem, the model training module 20 in the embedded object recognition system 100 adopts a dynamic rolling convolution algorithm to generate cognitive model parameter components that can be directly compiled and used under the embedded engineering framework. In view of the controllability and decomposability of the embedded terminal to the image acquisition process, the system integrates the image acquisition process with the first convolutional layer of the network, and designs a dynamic rolling convolution algorithm.

The dynamic rolling convolution algorithm divides the fusion rolling convolution algorithm into the following steps according to the decomposability and controllability of the image acquisition process of the embedded terminal camera:

(1) Obtain the pixel point data of the first k+H rows of the image and store them in the corresponding two-dimensional array G[H+1][S]. k∈[1,T-H-1]. Wherein, the value of S is the size of a single row of collected data, and after the acquisition is completed, the receiving of image data should be suspended through the corresponding control interface.

(2) Use the convolution kernel A[H][H] to perform convolution operation with the data from the kth row to the k+H-1th row of G, and store the obtained feature layer array in the feature layer in order For row k of the array, this traditional method is consistent, as shown in Figure 6(a) and Figure 6(b).

(3) Exchange the members in the k+1th to k+Hth rows of the two-dimensional array G with the members in the kth row to the k+Hth row respectively. In order to save memory space, only one additional variable is defined here for Exchange, for the elements in the adjacent row k and row k+1, exchange them in sequence according to the column order until all the elements are exchanged, and discard the original data of the kth row. The specific exchange method adopted in this paper is as shown in Figure 6 (c) shown.

(4) Turn on the read enable of the cache chip, continue to read the image data of row k+H+1, and store it in row k+H of G.

(5) At this time, increase the value of k by 1, and repeat steps (1) (2) (3) (4) in sequence until the convolution is completed and a complete output array is obtained.

In terms of space resource consumption, the input image storage space required by the fusion convolution algorithm of the embodiment of the present invention becomes S×(H+1), while the space occupied by the traditional convolution method is S×T.

The specific code process of the fusion rolling convolution algorithm of the embodiment of the present invention is shown in Table 3:

.

The embedded object recognition system based on image processing provided by the embodiments of the present invention can collect images in real time and obtain recognition results after reasoning through a lightweight convolutional neural network model in the model training module. Furthermore, the model training module in the system adopts the fusion rolling convolution algorithm to effectively optimize the embedded system's demand for the size and space of the image area. Furthermore, the system uses an optimized camera driving algorithm to drive the camera during the image acquisition process, which effectively improves the speed of image reading and display. The system can greatly reduce the demand for hardware resources while ensuring the accuracy of object recognition and reasoning speed, and realize the recognition and classification of different types of objects.

It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Accordingly, the embodiments should be regarded in all points of view as exemplary and not restrictive, the scope of the invention being defined by the appended claims rather than the foregoing description, and it is therefore intended that the scope of the invention be defined by the appended claims rather than by the foregoing description. All changes within the meaning and range of equivalents of the elements are embraced in the present invention. Any reference sign in a claim should not be construed as limiting the claim concerned.

In addition, it should be understood that although this specification is described according to implementation modes, not each implementation mode only includes an independent technical solution, and this description in the specification is only for clarity, and those skilled in the art should take the specification as a whole , the technical solutions in the various embodiments can also be properly combined to form other implementations that can be understood by those skilled in the art.

Claims

An embedded object recognition system based on image processing, characterized in that it includes: an image acquisition module for collecting image features of training objects; wherein, the image acquisition module uses an optimized camera driving algorithm to drive the camera to collect training Image features of the object; the model training module is connected with the image acquisition module, and the image features obtained by the image acquisition module are used as training materials, and the fusion rolling convolution algorithm is used to generate directly compiled and compiled under the embedded engineering framework. Cognitive model parameter component used; model terminal deployment module, used to deploy the cognitive model parameter component obtained by the model training module on the embedded terminal; terminal reasoning module, based on the target object collected by the image collection module image and use the cognitive model parameter component provided by the model terminal deployment module to recognize the target object.
An embedded object recognition system based on image processing according to claim 1, wherein the model training module includes two modes: a PC model training mode and an embedded terminal real-time reasoning training mode.
A kind of embedded object recognition system based on image processing according to claim 2, it is characterized in that, described PC model training mode comprises the steps: embedded terminal obtains image feature and image feature is transmitted to PC end; PC end Establish a data set according to the image features and generate a cognitive model parameter component that can be directly compiled and used under the embedded engineering framework according to the preset algorithm in the model training module; the cognitive model parameter component is burned. Deploy to embedded terminals.
An embedded object recognition system based on image processing according to claim 2, wherein the real-time reasoning training mode of the embedded terminal comprises the steps of: the embedded terminal obtains image features; the terminal reasoning module according to the image The feature performs image processing and generates cognitive model parameter components that can be directly compiled and used under the embedded engineering framework according to the preset algorithm in the model training module; the cognitive model parameter components are stored in the embedded terminal.
An embedded object recognition system based on image processing according to claim 1, wherein the hardware configuration in the model terminal deployment module is different according to the nature of parameters in the embedded object recognition system.
An embedded object recognition system based on image processing according to claim 5, wherein the constant parameters are stored in the FLASH memory, and the variable parameters are stored in the RAM memory.
A kind of embedded object cognition system based on image processing according to claim 6, it is characterized in that, described constant parameter comprises filter to eat away, offset BIAS parameter, propagation structure function; Described variable parameter comprises image characteristic, Input variables and output variables.
The embedded object recognition system based on image processing according to claim 6, wherein the embedded object recognition system erases and reads and writes the continuous address space of the preset FLASH designated sector Replaces reading and writing of dynamic arrays in RAM.
An embedded object recognition system based on image processing according to claim 1, wherein the parameter format used in the preset algorithm in the model training module is a multidimensional array form in C language.
A kind of embedded object cognition system based on image processing according to claim 1, is characterized in that, comprises image processing unit in the described image acquisition module, and described image processing unit adopts threshold filtering method to describe image acquisition module The obtained original image of the training object is processed to obtain the image features of the training object.