CN107978110A - Fence intelligence identifying system in place and recognition methods based on images match - Google Patents

Abstract

The invention discloses an image matching-based intelligent on-site recognition system and a recognition method for electronic fences. The system includes front-end information acquisition equipment, image acquisition card and upper computer. The front-end information acquisition equipment includes a CCD camera and a light source, connected to an image acquisition card, and the image acquisition card is connected to a host computer. The upper computer includes a client, an image processing module and a decision output module. The client is connected to the image processing module and connected to the CCD, and the image processing module outputs to the decision output module. This system is based on Hu invariant moments and improved BP neural network, and recognizes objects in situ through morphological feature matching. The image processing module performs electronic fence determination, image processing, extraction of Hu invariant moments, and training of BP neural network, etc., and the trained and mature BP classifier performs in-situ classification and recognition of objects, and outputs the results to the decision output module. The system has a simple structure, is suitable for most scenarios, and can quickly identify objects in the electronic fence.

Description

Intelligent on-site identification system and identification method for electronic fence based on image matching

technical field

The invention relates to the field of electronic fence monitoring and computer image processing, in particular to an image-matching-based intelligent on-site recognition system and a recognition method for an electronic fence, which is suitable for object on-site classification and recognition in most scenarios.

Background technique

With the advancement of social science and technology, human production behaviors and lifestyles are also constantly developing and changing, new technologies and new products are constantly emerging, and a large number of life and production activities occur, making intelligent monitoring and identification systems increasingly required by people. The real-time monitoring of a specified area to ensure on-site safety or identify various activities in the monitored area is inseparable from the support of electronic fence technology. Traditional and common electronic fences generally use photoelectric sensors to form an infrared light curtain, and detect whether there is an object intruding into the fence through the occlusion and reflection of infrared light, thereby triggering an alarm or taking pictures. However, this traditional electronic fence needs to provide power and lay electronic boundaries on site, and the expenditure of financial, material and manpower is relatively large.

The target recognition system is an important part of the current and future weapon system, and it is also an indispensable technical component of national production and social development. In modern industrial automation production, military targets such as aircraft, missiles, ships, etc., as small as auto parts, electronic assembly lines, electronic components, etc., involve a variety of inspection, measurement and target recognition applications. Traditional object recognition systems face the following problems:

(1) Images with a large amount of data, such as the detection and recognition of the size and shape of parts in the mass production process, the matching calculation of the system is very large;

(2) The displacement, rotation, and scale changes of the target object have different position changes relative to the sensor, and it is difficult to fix the shooting angle and environment. Many systems can no longer correctly identify the target when the target object changes in displacement, rotation, and scale;

(3) Applications ranging from small residential area monitoring, workshop production sites to mass production or assembly lines, and even in the military field and aerospace product research and development process, most of them require the system to be able to monitor, identify and give early warning online in real time, which requires the system to have Very fast processing speed and recognition speed.

Contents of the invention

In order to overcome the defects of the above-mentioned technologies, the present invention proposes an intelligent on-site recognition system of electronic fence based on image matching. Real-time rapid classification and recognition of objects.

To achieve the above object, the technical scheme of the present invention is as follows:

An image matching-based electronic fence intelligent on-site recognition system and recognition method are suitable for classification and recognition of various targets in various scenarios, and are especially suitable for online recognition of mass-produced parts or target objects on an assembly line. The electronic fence recognition system includes a front-end detection device 1 , an image acquisition card 2 and a host computer 3 .

The host includes a client, an image processing module and a decision output module. The client is connected to the front-end detection equipment and the image processing module to control the start of the CCD camera, the calibration of the electronic fence and other human-computer interaction work; the image processing module completes scene recognition, image processing, invariant moment extraction and BP neural network training and On-site recognition of the target; the decision output module receives and displays the recognition results from the image processing module.

The electronic fence area is determined by a scene with a specific mark drawn, and there is a pre-established scene library in the host computer, including a scene template with a specific mark drawn. The invention stipulates that the scene identification is simple and easy to identify. The scene identification templates in the scene library are some simple color blocks, characters, dashed lines, etc., such as two-dimensional codes, the word "workshop", playground runways and parking spaces, etc. The computer performs scene recognition through simple methods such as two-dimensional code scanning, template matching, and character recognition, and the client controls the image processing module to draw the boundary on the scene image to determine the electronic fence area.

For the first time, the present invention adopts the method based on fusion of Hu invariant moments and improved BP neural network model in the electronic fence recognition system, and performs object in-situ recognition through shape feature matching, and realizes that the target object in the electronic fence does not change with displacement, rotation, scale, etc. The accurate and fast recognition of changes can also realize the classification and recognition of various target objects.

An image-matching-based intelligent on-site identification system and identification method for electronic fences, comprising the following steps:

1) When the system is working, the host computer client is operated to start the CCD camera to capture images, and the images are decoded by the image acquisition card and sent to the host computer image processing module;

2) Start the system for the first time. After the image processing module receives the scene image, it first matches with the pre-established scene library in the host computer, and uses template matching, QR code scanning or character recognition to identify the specific scene area to determine the electronic image. fenced area;

3) After matching and detecting the specific scene area logo, the host computer quickly locates the area as the electronic fence area, and the client draws the boundary line on the scene image to define the electronic fence boundary through the image processing software, and then takes the image of the same area , delineated according to the geo-fence boundary of the first image;

4) After the host computer image processing module completes the calibration of the electronic fence, the following steps are performed on the image:

a) Preprocessing: Preprocessing the image collected by the CCD industrial camera, smoothing and filtering the grayscale image to overcome noise interference.

b) Image segmentation: image segmentation is performed by edge detection or threshold segmentation, and the foreground of the image is separated, that is, the target object and the background.

c) feature extraction: the intelligent on-site recognition system and recognition method for electronic fences based on image matching are characterized in that they are identified through shape matching of objects, so it is necessary to analyze the shape of the foreground image and extract shape features. Shape-based descriptions are divided into contour-based shape descriptions and region-based shape descriptions. The present invention adopts the latter, and adopts seven region Hu invariant moments based on shapes that do not change with displacement, rotation, and scale transformation for identification.

T ₁ =N ₂₀ +N ₀₂

T ₂ =(N ₂₀ -N ₀₂ ) ² +4N ₁₁ ²

T ₃ ＝(N ₃₀ -3N ₁₂ ) ² +(3N ₂₁ -N ₀₃ ) ²

T ₄ =(N ₃₀ +N ₁₂ ) ² +(N ₂₁ +N ₀₃ ) ²

T ₅ ＝(N ₀₃ -3N ₁₂ )(N ₃₀ +N ₁₂ )[(N ₃₀ +N ₁₂ ) ² -3(N ₂₁ +N ₀₃ ) ² ]

+(3N ₁₂ -N ₀₃ )(N ₁₂ +N ₀₃ )[3(N ₃₀ +N ₁₂ ) ² -(N ₂₁ +N ₀₃ ) ² ]

T ₆ ＝(N ₂₀ -N ₀₂ )[(N ₃₀ +N ₁₂ ) ² -(N ₂₁ +N ₀₃ ) ² ]+4N ₁₁ (N ₃₀ +

N ₁₂ )(N ₁₂ +N ₀₃ )

T ₇ ＝(3N ₂₁ -N ₀₃ )(N ₃₀ +N ₁₂ )[(N ₃₀ +N ₁₂ ) ² -3(N ₂₁ +N ₀₃ ) ² ]

+(3N ₁₂ -N ₀₃ )(N ₁₂ +N ₀₃ )[3(N ₃₀ +N ₁₂ ) ² -(N ₂₁ +N ₀₃ ) ² ]

where N _pq is the (p+q) order normalized center distance. Extract 7 invariant moments T ₁ -T ₇ , and normalize the 7 invariant moments, and use the normalized eigenvalues as the input vector of the BP neural network in step 7).

5) The upper computer establishes a neural network, first judges whether the system has completed the training of the BP neural network to make it reach the required recognition accuracy, if not, then proceeds to step 6), if there is a mature BP classification recognizer for training, Then proceed to step 8);

6) Set the parameters of the BP neural network, repeat step 1) to collect a number of sample images (such as 5000 pieces) of the target object;

7) Perform steps 2) to 4) on the sample image to extract core features, transfer the sample shape invariant moment feature vector into the BP neural network for training, and perform network connection weights according to the sample image invariant moment vector Adjust until the output error of the network is less than a certain value;

8) Perform morphological recognition on the collected images with the trained BP classification recognizer, and judge whether the target object appears in the fence through feature matching;

9) If the BP classifier recognizes that there is a target object in the electronic fence, the decision output device will give the recognition result and display the image of the target object.

In step 5), an improved BP neural network is established. The BP neural network is a multi-layer feed-forward network trained by error backpropagation. Its basic idea is the gradient descent method, and the actual output value of the network and the expected The mean square error of the output value is the smallest. The invention adopts an additional momentum method to improve the BP network model, and transmits the last weight (or threshold) change through a momentum factor. When the momentum factor takes a value of 0, the weight (or threshold) change is generated according to the gradient descent method. When the momentum factor takes a value of 1, the new weight (or threshold) change is set as the last weight (or threshold) change. The weight adjustment formula with additional momentum factor is:

Δw _ij (k+1)=(1-mc)ηδ _i P _j +mcΔw _ij (k)

Δb _i (k+1)=(1-mc)ηδ _i +mcΔb _i (k)

In the formula, Δw _ij represents the connection weight change between the i-th node of the hidden layer and the j-th node of the input layer; k is the number of training; η is the learning rate; δ _i represents the local minimum of the weight value approaching the error surface The value approach degree of , P _j represents the gradient; Δb _i is the threshold change of the i-th neuron; mc is the momentum factor, generally around 0.95. The momentum term added by this method reduces the shock in training, and at the same time increases the learning rate and improves the convergence.

Setting the node parameters of the BP neural network can realize the classification and rapid identification of different target objects. Step 6) in the parameter setting, the number of input layer nodes is determined according to the number of feature vectors, the number of output layer nodes is determined according to the output results required by the customer, and the number of hidden layer nodes is determined by the empirical formula Sure. Among them, m is the number of hidden layer nodes, l and n are the number of input and output nodes respectively, and α is a constant between 1 and 10.

The BP classifier is divided into two stages: training and learning and job recognition. In the former stage, the connection weights of the network are adjusted according to the sample images until the output error of the network is less than a certain value (step 7); in the latter stage, the recognition is step 8 ), only forward calculation without error backpropagation, to achieve in-situ recognition of objects in the electronic fence.

Compared with the prior art, the system has the following advantages:

The image processing software of the upper computer is used to recognize the scene of the area with a specific logo, and the client side calibrates the image boundary to establish a virtual electronic fence. It is not necessary to lay an electronic fence on the site, saving financial, material, and human resources.

For the first time, the system adopts the method of integrating Hu invariant moments and BP neural network in the electronic fence system, and uses Hu invariant moments to represent the shape characteristics of objects, which can realize the in-situ recognition of target objects that do not change with translation, rotation, scale transformation, etc. .

The system adopts the improved BP neural network model. Through the learning and training of a large number of sample images, the recognition speed is fast, the calculation amount is small, and the manual intervention is small. Improve the accuracy and speed of recognition.

Description of drawings

Fig. 1 is the flowchart of electronic fence recognition system of the present invention;

Fig. 2 is a system diagram according to an embodiment of the present invention;

Fig. 3 is the structural diagram of BP neural network algorithm;

Fig. 4 is the flow chart of BP neural network algorithm;

Fig. 5 is a flow chart of Hu invariant moment extraction.

Detailed ways

Embodiments of the present invention will be described below in conjunction with the accompanying drawings.

Figure 1 is a flow chart of the electronic fence identification system of the present invention, and Figure 2 is a system diagram according to an embodiment of the present invention, an image matching-based intelligent on-site identification system and identification method for electronic fences, applicable to various scenarios A variety of target recognition, especially suitable for mass-produced parts or target object recognition on the assembly line. The electronic fence is a virtual electronic fence, and the electronic fence identification system includes a front-end detection device 1 , an image acquisition card 2 and a host computer 3 .

The front-end detection equipment includes a CCD camera and a light source. The surface to be tested is a wide, continuous surface with a single color, which requires continuous, online, and high-speed detection. Therefore, a linear array black and white CCD and a line light source are used. This embodiment selects Canada DALSA Coreco company to produce, the model is the black-and-white linear array CCD camera of P2-22-04K30 and the LSL-450-34-R red light LED convergence line light source produced by Dongguan Kangshida Automation Technology Co., Ltd., reducing Data processing capacity, improve detection speed. The front-end detection equipment is installed in the area to be detected to take on-site images.

The image acquisition card receives the image decoding from the CCD camera and transmits it to the host computer. The selection of the image acquisition card depends on the interface of the CCD camera selected. If the selected camera has a USB interface, etc., it can be omitted. According to the selected CCD, this embodiment uses the DALSA Xcelera-CL PX4 Dual image acquisition card to receive the image from the CCD camera, decode it and transmit it to the host computer.

The host computer includes a client, an image processing module and a decision output module. The client is connected to the front-end detection equipment and image processing module to control human-computer interaction such as the start of the CCD camera and the calibration of the electronic fence; the image processing module completes scene recognition, image processing, BP classifier training and recognition; the decision output module receives information from The recognition result of the image processing module is displayed.

The electronic fence area is determined by a scene with a specific mark drawn on it. The scene in this embodiment is defined as a production workshop, and the specific mark is defined as the word "workshop" (or the texture of the conveyor belt). The quantity and type of products on the production line of the workshop can be detected online, and defect detection can also be performed through feature extraction and BP neural network parameter training.

The identification method of this electronic fence identification system is the first time in the electronic fence system to adopt the method based on the fusion of Hu invariant moments and BP neural network, and train a BP classifier with automatic identification function.

As shown in Figure 1, the method of the embodiment of the present invention comprises the following steps:

1) When the system is working, the host computer client is operated to start the CCD camera to capture images, and the images are decoded by the image acquisition card and sent to the host computer image processing module;

2) Start the system for the first time. After the image processing module receives the scene image, it first matches it with the pre-established scene library in the host computer. The scene in this embodiment is a workshop, and the scene identification is defined as the word "workshop" (or conveyor belt texture Also can), carry out character recognition to image and detect " workshop " two characters and carry out scene recognition.

3) After matching and detecting the specific scene area logo, the upper computer quickly locates the area as the electronic fence area, and the client uses the image processing software to draw the boundary line on the workshop site image to define the electronic fence boundary, and then the same area is photographed Image, defined according to the geo-fence boundary of the first image;

4) After the host computer image processing module completes the calibration of the electronic fence, the following steps are performed on the image:

a) Preprocess the image and use Wiener filter to denoise: use Wiener filter to remove noise and deblurring the image, the idea is to minimize the mean square error between the original image and the restored image. Immediate:

Where f(x,y) is the original image, To restore the image, E represents the mean square error of the two.

b) Image segmentation, the present embodiment adopts the method based on Sobel operator edge detection to perform image segmentation.

In the formula, A is the original image, G _x , G _y represent the image detected by horizontal and vertical edges respectively, G is the gradient size, and θ is the gradient direction. Sobel operator can directly calculate G _x , G _y to detect the presence of edges, and the change from dark to bright and from bright to dark. However, since only two-direction templates are used, it can only detect vertical and horizontal edges, which is not suitable for images with complex textures. The surface to be tested in this embodiment is a wide and continuous surface, and this embodiment is suitable for detecting objects with simple edges such as components and large objects.

c) As shown in Figure 5, it is a flow chart of feature extraction. An on-site recognition system for electronic fences based on image matching is identified through morphological matching, so shape features need to be extracted. Shape-based descriptions are divided into contour-based shape descriptions and region-based shape descriptions. The present invention adopts the latter, and the following seven shape-based region invariant moments that do not change with translation, rotation, and scale transformation are combined from the normalized second-order and third-order central moments.

T ₁ =N ₂₀ +N ₀₂

T ₂ =(N ₂₀ -N ₀₂ ) ² +4N ₁₁ ²

T ₃ ＝(N ₃₀ -3N ₁₂ ) ² +(3N ₂₁ -N ₀₃ ) ²

T ₄ =(N ₃₀ +N ₁₂ ) ² +(N ₂₁ +N ₀₃ ) ²

T ₅ ＝(N ₀₃ -3N ₁₂ )(N ₃₀ +N ₁₂ )[(N ₃₀ +N ₁₂ ) ² -3(N ₂₁ +N ₀₃ ) ² ]

+(3N ₁₂ -N ₀₃ )(N ₁₂ +N ₀₃ )[3(N ₃₀ +N ₁₂ ) ² -(N ₂₁ +N ₀₃ ) ² ]

T ₆ ＝(N ₂₀ -N ₀₂ )[(N ₃₀ +N ₁₂ ) ² -(N ₂₁ +N ₀₃ ) ² ]+4N ₁₁ (N ₃₀ +

N ₁₂ )(N ₁₂ +N ₀₃ )

T ₇ ＝(3N ₂₁ -N ₀₃ )(N ₃₀ +N ₁₂ )[(N ₃₀ +N ₁₂ ) ² -3(N ₂₁ +N ₀₃ ) ² ]

+(3N ₁₂ -N ₀₃ )(N ₁₂ +N ₀₃ )[3(N ₃₀ +N ₁₂ ) ² -(N ₂₁ +N ₀₃ ) ² ]

where N _pq is the (p+q) order normalized center distance. According to the above formula, there are only slight differences in the values under different translation, rotation, and scale changes, which can be attributed to the numerical calculation error of the discrete image. Thus, the embodiment of the present invention can realize the recognition of objects that do not change with translation, rotation, and scale transformation, extract 7 invariant moments T ₁ ~ T ₇ , and perform normalization processing on the 7 invariant moments, and normalize The eigenvalue after normalization is used as the input vector of the BP neural network in step 7).

5) The upper computer establishes a neural network, first judges whether the system has completed the training of the BP neural network to make it reach the required recognition accuracy, if not, then proceeds to step 6), if there is a mature BP classification recognizer for training, Then proceed to step 8);

6) Set various parameters of the BP neural network, the input node is 7, the middle layer can be selected between 4 and 13, and the output node is set to 2. Repeat step 1) to collect 5000 sample images of the target object;

7) Perform steps 2) to 4) on the sample image to extract core features, transfer the sample shape invariant moment feature vector into the BP neural network for training, and perform network connection weights according to the sample image invariant moment vector Adjust until the output error of the network is less than a certain value;

8) Perform morphological recognition on the collected images with the trained BP classification recognizer, and judge whether the target object appears in the fence through feature matching;

9) If the BP classifier recognizes that there is a target object in the electronic fence, the decision output device will give the recognition result and display the image of the target object.

Step 5) of this system establishes an improved BP neural network, as shown in Figure 3 is the structural diagram of the BP neural network algorithm, and Figure 4 is a flow chart of the three-layer BP neural network algorithm.

There are many defects in the traditional BP neural network algorithm, such as easy to fall into local minimum, slow learning convergence speed, etc. The present invention adopts the method of additional momentum to improve the BP network model. The additional momentum method adds a value proportional to the previous weight change to each weight change based on the backpropagation method, and generates a new weight change according to the backpropagation method. The activation function of the BP neural network is selected as the S-type activation function:

The weight adjustment formula with additional momentum factor is:

Δw _ij (k+1)=(1-mc)ηδ _i P _j +mcΔw _ij (k)

Δb _i (k+1)=(1-mc)ηδ _i +mcΔb _i (k)

In the formula, Δw _ij represents the connection weight change between the i-th node of the hidden layer and the j-th node of the input layer; k is the number of training; η is the learning rate; δ _i represents the local minimum of the weight value approaching the error surface The value approach degree of , P _j represents the gradient; _Δbi is the threshold value change of the i-th neuron; mc is the momentum factor, which is taken as 0.95. The momentum term added by this method reduces the shock in training, and at the same time increases the learning rate and improves the convergence.

The judgment conditions for using the momentum method in the training program design of the additional momentum method are:

Where E(k) is the sum of squared errors of the Kth step.

The design of BP network classifier considers the design of input layer, hidden layer and output layer. The more hidden layers, the higher the accuracy, but the higher the complexity. In this embodiment, the recognition object type is set to two kinds of parts on the production line, then the number of output nodes is 2, the number of neurons in the input layer is 7, and the number of input nodes is the number of core feature vectors 7, corresponding to 7 invariant moments T ₁ ~ T ₇ , the number of hidden layer nodes is determined by the empirical formula Sure. Among them, m is the number of hidden layer nodes, l and n are the number of input and output nodes respectively, and α is a constant between 1 and 10, so m can be a number between 4 and 13. Therefore, the node settings in step 6) are as follows: the input node is 7, the output node is 2, and the intermediate layer nodes can be optimally selected between 4 and 13.

The BP classifier is divided into two stages: training and learning and job recognition. In the former stage, the connection weights of the network are adjusted according to the sample images until the output error of the network is less than a certain value (step 7); in the latter stage, the recognition is step 8 ), only forward calculation without error backpropagation, to achieve in-situ recognition of objects in the electronic fence.

The decision output device and the human-computer interaction interface are built on the same operation display, and the results of system recognition can be seen in real time when the client is operating.

The above-mentioned embodiment illustrates the specific implementation of the present invention, but it is not the only embodiment of the present invention, and does not constitute a limitation to the protection scope of the present invention.

Claims (4)

1. A kind of 1. fence based on images match intelligently identifying system in place, including front-end information acquisition equipment (1) image Capture card (2) and host computer (3), it is characterised in that the front-end information, which obtains equipment (1), includes CCD industrial cameras and light Source, is placed in fence area；The image collected by front-end information acquisition equipment (1) is decoded through image pick-up card (2) Incoming host computer (3).
2. 2. the intelligent identifying system in place of the fence according to claim 1 based on images match, it is characterised in that institute The host computer stated includes client 1) image processing module 2) and decision-making output module 3), client 1) it is connected to image procossing mould Block 2), and be connected with front end detecting devices (1) CCD camera, image processing module 2) it is connected to decision-making output module 3)；Host computer Image processing module carries out scene Recognition to the region for being painted with specific identifier, and virtual electronic fence boundary is drawn by client；Figure As processing module also completes identification in place of pretreatment, image segmentation, feature extraction, the training of BP neural network and target etc. Function.
3. A kind of 3. fence of the intelligence of the fence based on the images match identifying system in place based on described in claim 1 Intelligence recognition methods in place, it is characterised in that comprise the following steps：

   1) when system works, by host computer client operation, CCD camera shooting image is started, image is via image pick-up card Decode and be passed to host computer image processing module；

   2) activation system first, after image processing module receives image scene, first by with being built up in advance in host computer Scene library is matched, and the methods of using template matches, two-dimensional code scanning or character recognition identifies special scenes region to determine Fence region；

   3) after matching detection is identified to specific scene areas, it is fence region that host computer, which positions the region, by client Draw boundary line on image at the scene by image processing software and define fence border, the figure of the same area shot afterwards Picture, delimited according to the fence border of first width image；

   4) after host computer image processing module completes the calibration of fence, the processing that is followed the steps below to image：

   4-1) pre-process：The image collected by CCD industrial cameras is pre-processed, gray level image is carried out at smothing filtering Manage to overcome noise jamming；

   4-2) image is split：Image segmentation is carried out with the methods of edge detection or Threshold segmentation, isolates the prospect i.e. mesh of image Mark object and background；

   4-3) feature extraction：Matched by the form of object to identify, it is necessary to foreground image progress shape analysis, extract shape Feature, the description based on shape are divided into the shape description based on profile and the shape description based on region, employ based on shape Not with displacement, rotation, change of scale and 7 region Hu changing not bending moment T₁~T₇It is identified,

   T₁=N₂₀+N₀₂

   T₂=(N₂₀-N₀₂)²+4N₁₁ ²

   T₃=(N₃₀-3N₁₂)²+(3N₂₁-N₀₃)²

   T₄=(N₃₀+N₁₂)²+(N₂₁+N₀₃)²

   T₅=(N₀₃-3N₁₂)(N₃₀+N₁₂)[(N₃₀+N₁₂)²-3(N₂₁+N₀₃)²]

   +(3N₁₂-N₀₃)(N₁₂+N₀₃)[3(N₃₀+N₁₂)²-(N₂₁+N₀₃)²]

   T₆=(N₂₀-N₀₂)[(N₃₀+N₁₂)²-(N₂₁+N₀₃)²]+4N₁₁(N₃₀+N₁₂)(N₁₂+N₀₃)

   T₇=(3N₂₁-N₀₃)(N₃₀+N₁₂)[(N₃₀+N₁₂)²-3(N₂₁+N₀₃)²]

   +(3N₁₂-N₀₃)(N₁₂+N₀₃)[3(N₃₀+N₁₂)²-(N₂₁+N₀₃)²]

   Wherein N_pqCentre-to-centre spacing is normalized for (p+q) rank.Extract 7 invariant moments T₁~T₇, and 7 invariant moments is normalized Processing, the input vector using the characteristic value after normalized as BP neural network in step 7)；

   5) host computer establishes a neutral net.First determine whether that the training whether system has been completed to BP neural network reaches it To desired accuracy of identification, if nothing, step 6) is transferred to, if the ripe BP Classification and Identification devices of training, carry out step 8)；

   6) parameters of BP neural network are set, and input layer is several to be determined according to feature vector number, output layer number of nodes The output result needed according to client determines that node in hidden layer m is by empirical equationDetermine, its Middle m is node in hidden layer, and l, n are respectively inputoutput section points, and α is the constant between 1 to 10；Repeat step 1) if collection The sample image of dry target object；

   7) step 2)~4 are carried out to sample image), core feature is extracted, by the incoming BP god of sample shape invariant moment features vector Through being trained in network to it, according to sample image, bending moment vector is not adjusted the connection weight of network, until network Output error be less than a certain numerical value；

   8) the BP Classification and Identifications device completed with training carries out form identification to the image collected, judges target by characteristic matching Whether object is appeared in fence；

   If 9) BP graders identify there is target object in fence, recognition result is provided by decision-making output device and is shown The image of target object.
4. 4. the intelligence of the fence based on the images match knowledge in place according to claim 3 based on described in claim 1 The fence of other system intelligently recognition methods in place, it is characterised in that the BP neural network described in step 5) is using additional Momentum method improved BP network, is shaken with reducing in training, speeds learning rate.Weights with the additional momentum factor are adjusted Formula is：

   Δw_ij(k+1)=(1-mc) η δ_iP_j+mcΔw_ij(k)

   Δb_i(k+1)=(1-mc) η δ_i+mcΔb_i(k)

   In formula, Δ w_ijRepresent that i-th of node of hidden layer and the connection weight of j-th of node of input layer change；K is frequency of training；η For learning rate；δ_iRepresent weights level off to error surface local minimum value convergence degree, P_jRepresent gradient；Δb_iFor i-th The changes of threshold of neuron；Mc is factor of momentum, takes 0.95.