CN108157219A - A kind of pet based on convolutional neural networks stops apparatus and method of barking - Google Patents

Abstract

The present invention provides a method for stopping barking of pets based on convolutional neural network, comprising the following steps: S1, preparing training samples, selecting several sections of pet barking sounds as the training data of the model; Bark sound preprocessing; S3, calculating the spectrogram; S4, inputting the spectrogram into the convolutional neural network; S5, training the model; S6, pet identification; S7, triggering the stop bark sound, and playing the stop bark sound. The present invention also provides a pet barking prevention device based on convolutional neural network. The beneficial effect of the present invention is that: applying the convolutional neural network to the method for stopping barking of pets improves the flexibility and anti-interference performance of stopping barking without causing harm to pets, and can also identify pets.

Description

A device and method for preventing barking of pets based on convolutional neural network

technical field

The present invention relates to a barking prevention method, in particular to a pet barking prevention device and method based on a convolutional neural network.

Background technique

Traditional anti-bark methods include the use of simple electronic anti-bark devices, such as vibrating, ultrasonic or electric shock types, in addition to surgically cutting vocal cords for pets (such as dogs) and wearing masks for pets.

The disadvantages of traditional methods are as follows:

(1) Inflexible, easy to cause harm to pets, etc.

(2) It is impossible to identify the pet that triggers the barking sound.

Contents of the invention

In order to solve the problems in the prior art, the present invention provides a pet barking prevention device and method based on a convolutional neural network.

The present invention provides a pet barking device based on convolutional neural network, comprising a microphone, an operational amplifier, an embedded processor, a memory, a power amplifier and a speaker, wherein the output of the microphone is connected to the input of the operational amplifier The output end of the operational amplifier is connected with the input end of the embedded processor, the embedded processor is connected with the memory, the output end of the embedded processor is connected with the input of the power amplifier The output terminal of the power amplifier is connected with the input terminal of the speaker.

The present invention also provides a method for stopping barking of pets based on convolutional neural network, comprising the following steps:

S1. Prepare training samples and select several segments of pet barking sounds as training data for the model;

S2, preprocessing, preprocessing the original pet barking sound;

S3, calculating the spectrogram;

S4, input the spectrogram into the convolutional neural network;

S5, training model;

S6. Pet identification;

S7, triggering the bark-stopping sound, and playing the bark-stopping sound.

As a further improvement of the present invention, the preprocessing in step S2 includes pre-emphasis, frame-by-frame windowing, and barking sound endpoint detection.

As a further improvement of the present invention, the convolutional neural network in step S4 includes a convolutional layer, a downsampling layer, and a fully connected layer; the convolutional layer, as the first layer of the convolutional neural network, directly performs a two-dimensional spectrogram signal Convolution operation; the size of the convolution kernel filter is selected as a 5X5 template, and the results obtained through different convolution kernel filters constitute a feature map; each convolution kernel filter shares the same parameters, including the same weight matrix and bias Setting items, the mathematical model of the convolutional layer is as follows:

y=f(x*k+b)

Among them, x is the input signal, k is the convolution kernel, * is the convolution operation, b is the bias term, f is the sigmoid function, and y is the output feature map;

The downsampling layer is deployed after the convolutional layer, the downsampling filter selects a 2X2 template, the sampling strategy takes the maximum value corresponding to 4 pixels, and the fully connected layer sends the score value to the classifier.

As a further improvement of the present invention, in step S5, the model training is completed on the computer PC, and the parameters are adjusted to optimize the training model through forward propagation and backward propagation. In order to better deploy the model trained on the PC to the embedded mobile terminal with relatively scarce computing resources, it is necessary to quantify the weights and simplify the model to generate an APK model file supported by Android.

As a further improvement of the present invention, in step S6, the trained apk model file is deployed in the pet-embedded Android anti-bark device based on convolutional neural network according to claim 1, and the microphone in the pet anti-bark device collects pet The barking sound signal is extracted from the sound spectrogram, and it is used as the input of the convolutional neural network model to obtain the score probability value. The score probability value is compared with the set threshold to judge. If it is greater than the threshold, the identity of the pet to be detected is confirmed, otherwise unconfirmed.

As a further improvement of the present invention, in step S7, after determining the pet's identity, it is detected whether the amplitude of the pet's barking sound exceeds the set threshold, and if it exceeds, the barking sound stored in the memory is played through the speaker.

The beneficial effect of the present invention is that: applying the convolutional neural network to the method for stopping barking of pets improves the flexibility and anti-interference performance of stopping barking without causing harm to pets, and can also identify pets.

Description of drawings

FIG. 1 is a schematic diagram of a pet barking prevention device based on a convolutional neural network according to the present invention.

Fig. 2 is a schematic flowchart of a method for stopping barking of pets based on a convolutional neural network according to the present invention.

Detailed ways

The present invention will be further described below in conjunction with the description of the drawings and specific embodiments.

As shown in Figure 1, a kind of pet anti-bark device based on convolutional neural network, comprises microphone 101, operational amplifier 102, embedded processor 103, memory 104, power amplifier 105 and loudspeaker 106, wherein, the microphone 101 The output terminal is connected with the input terminal of the operational amplifier 102, the output terminal of the operational amplifier 102 is connected with the input terminal of the embedded processor 103, the embedded processor 103 is connected with the memory 104, and the The output end of embedded processor 103 is connected with the input end of described power amplifier 105, and the output end of described power amplifier 105 is connected with the input end of described speaker 106, and this pet anti-barking device based on convolutional neural network is embedded Android mobile terminal anti-barking device.

As shown in Figure 1, the present invention provides a pet barking device based on convolutional neural network. The microphone 101 collects the pet barking sound and outputs it to the operational amplifier 102 after the system pre-processes the pet barking sound. After the signal of the operational amplifier 102 is amplified, it is input into the recognition model of the embedded processor 103. After the model recognizes the correct pet identity, the pet owner's barking sound recorded in advance by the memory 104 is played by the power amplifier 105 and the loudspeaker 106. , the pet stops barking after hearing the sound of the owner stopping barking, so as to achieve the purpose of stopping barking.

Convolutional neural network is currently the most researched and most successfully applied model in the deep learning system. It is widely used in many fields such as images, voice, and video, and has made great contributions in the field of artificial intelligence. The invention breaks through the traditional technical means, and uses the convolutional neural network in the sound recognition of bark stop. The method is divided according to functions, and mainly includes the training of the pet barking sound model, pet identification, and the triggering of the barking stop sound (generally referring to the barking stop sound recorded in advance by the pet owner). Among them, model training and model identification are completed on the computer and the embedded Android mobile terminal anti-barking device respectively.

As shown in Figure 2, a method for stopping barking of pets based on convolutional neural network includes the following steps:

1. Training of pet barking sound model

(1) Prepare training samples

Select 20 pieces of barking sounds from pets (such as pet dogs) as model training data, and the length of each sound is about 30 seconds.

(2) Pretreatment

In order to extract useful barking signal and reduce the impact of environmental noise, it needs to be preprocessed. The preprocessing methods adopted in this solution include pre-emphasis, frame division and windowing, barking sound endpoint detection, etc.

(3) Calculate spectrogram

Considering the pet barking sound information, the spectrogram is used here as the input of the convolutional neural network. The spectrogram contains a large amount of information related to the characteristics of the barking sound, and it combines the advantages of the spectrogram and the time-domain waveform.

(4) Convolutional neural network

A typical convolutional neural network structure is used here, and its structure includes convolutional layers, downsampling layers, and fully connected layers. As the first layer of the convolutional neural network, the convolution layer directly performs convolution operations on the two-dimensional spectrogram signal. Convolution and filter size selection 5X5 template. The results obtained through different convolution kernels constitute the feature map. Each convolution kernel filter shares the same parameters, including the same weight matrix and bias term. The mathematical model of the convolutional layer used here is as follows:

y=f(x*k+b)

Among them, x is the input signal, k is the convolution kernel, * is the convolution operation, b is the bias term, f is the sigmoid function, and y is the output feature map.

In order to increase the robustness of the system and reduce the computational complexity, the input is down-sampled, and the down-sampling layer is deployed after the convolutional layer. The downsampling filter selects a 2X2 template, and the sampling strategy takes the maximum value corresponding to 4 pixels. The fully connected layer sends the score value to a classifier (such as a softmax classifier).

(5) Training model

The model training is completed on the computer PC. Through forward propagation and backward propagation, the parameters are adjusted to make the training model optimal. In order to better deploy the model trained on the PC to the embedded mobile terminal with relatively scarce computing resources, it is necessary to quantify the weights and simplify the model to generate an APK model file supported by Android.

2. Pet identification

The trained APK model is deployed in the embedded Android mobile device. The microphone 101 in the embedded Android mobile terminal barking device collects the pet barking sound signal, extracts the spectrogram signal volume, uses the spectrogram as the input of the convolutional neural network model, and obtains a score probability value, which is greater than the set value. Threshold, the pet to be detected is confirmed, otherwise it is not confirmed.

3. Triggered by barking sound

After determining the pet's identity, it is detected whether the amplitude of the pet's barking sound exceeds the set threshold, and if it exceeds, the barking sound stored in the memory 104 will be played through the loudspeaker 106 .

The present invention provides a device and method for preventing barking of pets based on convolutional neural network, which creatively applies convolutional neural network to pet barking prevention, improves the flexibility and anti-interference performance of barking suppression, and does not affect The pet causes harm, and the present invention can also identify the pet in addition.

A convolutional neural network-based pet barking device and method provided by the present invention has the following characteristics:

1. Human voice recognition technology is applied to pet barking recognition.

2. The convolutional neural network is used as a model for training and recognition.

3. Extract the spectrogram features of pet barking sounds as the input of the convolutional neural network.

The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be assumed that the specific implementation of the present invention is limited to these descriptions. For those of ordinary skill in the technical field of the present invention, without departing from the concept of the present invention, some simple deduction or replacement can be made, which should be regarded as belonging to the protection scope of the present invention.

Claims (7)

1. The device 1. a kind of pet based on convolutional neural networks only barks, it is characterised in that：Including microphone, operational amplifier, embedding Enter formula processor, memory, power amplifier and loudspeaker, wherein, the output terminal of the microphone and the operational amplifier Input terminal connects, and the output terminal of the operational amplifier is connect with the input terminal of the embeded processor, the embedded place Reason device is connect with the memory, and the output terminal of the embeded processor is connect with the input terminal of the power amplifier, institute The output terminal for stating power amplifier is connect with the input terminal of the loudspeaker.
2. A kind of method 2. pet based on convolutional neural networks only barks, which is characterized in that include the following steps：

   S1, prepare training sample, several sections of pets is selected to yelp training data of the sound as model；

   S2, pretreatment, the sound that yelps to original pet pre-process；

   S3, sound spectrograph is calculated；

   S4, sound spectrograph is inputted into convolutional neural networks；

   S5, training pattern；

   S6, pet identification；

   S7, sounds trigger of only barking, play and stop bark sound.
3. The method 3. pet according to claim 2 based on convolutional neural networks only barks, it is characterised in that：In step S2 Pretreatment includes that preemphasis, framing adding window, yelp sound end-point detection.
4. The method 4. pet according to claim 2 based on convolutional neural networks only barks, it is characterised in that：In step S4 Convolutional neural networks include convolutional layer, down-sampled layer and full articulamentum；First layer of the convolutional layer as convolutional neural networks, directly Convolution operation is carried out to two-dimentional sound spectrograph signal；Convolution kernel wave filter size selects 5X5 templates, is filtered by different convolution kernels The result that device acts on constitutes characteristic pattern；Each convolution kernel wave filter shares identical parameter, including identical weight square Battle array and bias term, the convolutional layer mathematical model of use are as follows：

   Y=f (x*k+b)

   Wherein, x is input signal, and k is convolution kernel, and * is convolution operation, and b is bias term, and f is sigmoid functions, and y is that output is special Sign figure；

   Down-sampled layer is deployed in after convolutional layer, and desampling fir filter selection 2X2 templates, sampling policy takes 4 pixels corresponding Maximum value, full articulamentum give score value to grader.
5. The method 5. pet according to claim 2 based on convolutional neural networks only barks, it is characterised in that：In step S5 In, after model training and optimization, the APK model files of generation Android supports.
6. The method 6. pet according to claim 5 based on convolutional neural networks only barks, it is characterised in that：In step S6 In, the APK model files after training are disposed the pet described in claim 1 based on convolutional neural networks and are stopped in device of barking, The pet microphone acquisition pet in device that only barks yelps signal, sound spectrograph semaphore is extracted, using sound spectrograph as convolutional Neural The input of network model obtains scoring probability value, which is more than threshold value, and pet identity to be detected is confirmed, is not otherwise obtained Confirm.
7. The method 7. pet according to claim 6 based on convolutional neural networks only barks, it is characterised in that：In step S7 In, it is determined that it after pet identity, then detects pet and whether yelps amplitude more than the threshold value set, if being more than, deposited being stored in The only bark sound of reservoir, is played back by loudspeaker.