Music recognition algorithm based on T-S cognitive neural network

3.1 CNN structure based on T-S algorithm

Convolution neural network (convolutional neural network, CNN) because of its unique structure, at the same time of image feature extraction, will also be able to extract more details of image information. This not only solves the problem of many parameters and slow training in most traditional neural networks, but also prevents the occurrence of over fitting. Since AlphaGo defeated world go champion lee sedol in 2016, convolutional neural network has been pushed into a wave again, especially in the field of computer vision. Convolutional neural networks have two important characteristics: 1) Shared weights. In the traditional neural network, the weight w of each layer is only used once, and different weight w will be generated when it is used again. However, in the convolutional network, the convolution kernel is convolved with each pixel value (input vector) in the image, so only a set of weights is required. When the input vector is finished with the set of weights, it indicates that the convolution operation is completed. The design of Shared weights does not reduce the time consumption in the forward propagation stage, but to some extent, the number of weight parameters required by the whole model is greatly reduced, which greatly improves the computing performance of the computer. In the operation of convolution, the convolution kernel slides from left to right, from top to bottom on the input image in accordance with the given step size s until the end of the operation. Compared with the traditional neural network, the number of parameters of convolutional neural network is not only reduced, but also its operation speed is improved to a certain extent. 2) sparse connection. In order to mine the information of local association in image space, convolutional neural network adopts the mode of local connection by strengthening the nodes between adjacent layers in the neural network, and abandons the mode of full connection, that is, it adopts the mode of less than the input kernel to complete. For example, if there is a m a n input output, the traditional neural network to each output with each input matrix multiplication, the time complexity, and extraction only meaningful k convolutional neural network input, its time complexity is, because in actual application, general is far less than m, k is more practical significance, and this reduces the time complexity, on the one hand and improve the efficiency of storage.

The CNN algorithm is a decision behavior that uses fuzzy reasoning to imitate human being in the uncertain environment [6]. The

fuzzy rules are constantly adjusted to imitate the system output from the self-training function of the neural network from the initial given fuzzy rules [7, 8, 9]. In the fuzzy algorithm, two main methods are used to express the fuzzy set: one is expressed output set of the fuzzy rules, such as NB, PB, and so on. The other is the formula expression of the fuzzy rule after the input language variable, and the typical case is the linear combination of the input variables. Because the model is first proposed by Takagi and Sugeno, As a result, the T-S algorithm is usually called a fuzzy system.

The convolutional layer is the core structure of the network. Each neuron in this layer is locally connected with the previous layer, and the weight matrix connected is called convolution kernel or filter. The convolution kernel extracts the features of different positions of the input image in the way of “sliding window” with step size s, and outputs a feature graph. A feature graph corresponds to a convolution kernel, that is, the weights of each neuron in the feature graph are Shared. But the features extracted from different feature maps are different. The convolution kernel is the receptive field in the analog visual system, and the size of the convolution kernel

Corresponding to the magnitude of the sensory field, the direction of the convolution kernel corresponds to the direction of the sensory field axon. Different directions of the convolution kernel are used to extract features in different directions, and the first convolution layer is used to extract simple features such as edges or lines. Since the size of the output feature graph decreases with the depth of the convolutional neural network

In the transmission process, the image feature information is not missing, and the number of output feature graph should be increased layer by layer, so that the information in the transmission (the size of output feature graph multiplied by the number) is non-decreasing. Although only simple features are extracted in the low level of the network, with the increase of the network level, each neuron in the high level will feel an increasing number of low-level regions, that is, those simple features extracted in the low level will constantly converge to form complex features in the high level, and finally

obtain the global feature information

Each component xi of the input vector X is a fuzzy set. The set of linguistic variable values is T(xi)={Ai1,Ai2⋯Ami},   i=1,2,⋯n,where Aisi(si=1,2⋯mi)is si th language variable value of xi . Membership function:

If the output vector is Y, the fuzzy rules of the proposed methods is in the form of (1) Rj:F x1 is  A1s1j,x2 is  A2s2j,⋯,xn is  Ansji

where j=1,2,⋯m;m≤∏i=1mi.If the input quantity is fuzzed by the single point fuzzy set, the applicability of each rule can be aj=uA1s1j(x1)  ∧uA2s2j(x2)∧⋯∧uAnsji(xn)the output of the fuzzy system can be written as

According to the fuzzy model given above, system diagram of the proposed algorithm is shown in Figure 2. The network is composed of two parts of the forward part network and the post part network. The forward part network is satisfied with the fuzzy rules, and the post part is related to produce the post of the fuzzy rules.

Figure 2

The system diagram of the new proposed algorithm

We will analyze each layer of the network, and give the node functions of each layer:

Input layer: Each node is directly connected to each node xi , and it plays the role of sending the input information to the next level.

The input value of the zero nodes in the input layer is

The middle 1-layer of forward part network: Each node represents the value of one language variables. Its effect is to compute membership functions of all input components in fuzzy sets of linguistic variables.

uisi(xi)       whereuisi(i=1,2⋯n;si=1,2⋯mi). m is the input’ dimensions. mi is the fuzzy division number of si .

(3) The middle 2-layer of forward part network: Its main task is to compute the fitness of each rule.

(4) The middle 3-layer of the forward part

network: The implementation of this level is the normalization calculation.

(5) The middle layer of the post part network: Each node represents a rule whose action is to express the consequent of formulating rule.

(6) Output layer computing system:

It can be seen that Y is the cumulating of each rule post, and the weighting coefficient is the normalized applicability. Assuming that the number of fuzzy partitions of each input component is predetermined, the parameter Wji to learn is mainly the connection power of the post network. And the central values Ci,s_i and widths ei,s_i of the membership functions in the middle-layer of the front part are 1. Assuming the error cost function is E=12[Y¯−Y],  Y¯,Ymeans the expected output and real output.

3.2 Training process of connective parameters based on GA algorithm

In terms of the overall recognition effect, the recognition accuracy of the model trained in this paper is better than most popular methods. Because the braille music picture in this paper is based on the single-sided braille picture, the contrast here is only limited to the recognition and contrast of the single-sided braille picture. The CNN model is compared with feed forward neural network, BP neural network, fuzzy classification algorithm and standard distance positioning method used in recent years, which shows that the recognition accuracy of braille music image in this paper is higher than other methods.

Genetic algorithm is a iterative process. A group of candidate solutions are retained in iterative process. Some iterative operators, such as crossover and mutation, are used to generate a new generation of candidate solutions, which can be repeated to meet some convergence conditions. Genetic algorithms have better robustness. The main feature of genetic algorithms is that genetic algorithms use the encoding of the parameter set rather than the parameters itself; the genetic algorithm aims to guide the genetic process, not to seek derivatives or other theorems; the genetic algorithm searches for optimization in the point group rather than a point; the genetic algorithm uses the possibility transformation. Law, not the finalized law

The training of weights is a complex continuous parameter optimization problem. If binary coding is used, the encoding string will be too long and need to be decoded as real number to make the weight change step into step, thus affecting the accuracy of network learning. The brief block diagram of genetic algorithm is shown in the Figure 3. The optimal weight parameters can be obtained through the flow chart above.

Figure 3

The brief block diagram of genetic algorithm

4.1 Simulation experiment 1

The simulation platform is matlab7.0, and audio samples are from the shocking Music software released by Tsinghua University press in 1997. The sampling rate is 11025 Hz, and the framing standard of the original signal is divided into Hamming windows, 256 frames per frame and 128 frames. In the mixed signal, the voice sample is 42, and the music sample is 67, of which the music sample contains five categories, such as string music, percussion music, woodwind music, brass music and keyboard music features.

At this stage, in order to verify the effectiveness of the convolution neural network recognition model established in the experiment, the initialization of weight parameters in the experiment adopts random Numbers. Secondly, in order to make the model not fall into saturation prematurely during training and affect the learning ability of convolutional neural network, the random number used in the experiment will be relatively small. In this experiment, the training and testing were carried out in a batch-iteration manner, with 200 training iterations respectively, and the accuracy of test data was output every 20 times. Through training and testing, the variation trend of the accuracy is shown in figure 13. Through analysis, it can be seen that after 140 iterations, the accuracy of the model tends to be stable and close to 1. This is because: the structure of music symbol is relatively simple (with dot as the main feature),

and the image resolution used in training and testing is higher, that is, the image quality is better; Secondly, the recognition model is easy to extract the features of braille music symbols in braille music pictures, so that music symbols can be recognized quickly.

The recognition algorithm proposed in this paper compared with the conventional K-means clustering algorithm and short-time energy. The method of fixed clustering center is used in the simulation experiment to fix the clustering center of the mean and variance of the entropy of the phonetic and music samples. The input parameter is the mean and variance of the sample entropy value of each sample, and the output is the class of the sample. The simulation results are compared with the traditional methods of extracting short-time energy and MFCC as features. The results have shown as the above table 1.

4.2 Simulation experiment 2

The recognition process is based on the various states that are estimated by the feature flow, and the optimal state sequence is obtained to get the sequence of notes. In view of multiple candidates, the search process should give different length candidates, corresponding to different note detection results. Finally, the recognition results are obtained by multiple candidate decisions.

Recognizing a piece of piano music with 6 notes, the experiment shows that the system can recognize 55 notes correctly; 3 errors are all two consecutive notes with high homophone recognition as one. From hearing, the type error can be tolerated and the recognition rate is 90.2%. By controlling the parameters of the recognition module, this type of error can be avoided, and the recognition rate can be further improved.

For three pieces of music played by piano, Violin and Oboe, the accuracy rate of the new proposed algorithm is shown in the table 2.

As is shown in the Table 2 and Figure 4, the training set is the 61 note played for the piano. It can be seen from this table that the recognition rate is insensitive to both within and outside. The recognition error of piano music is mentioned in the above experimental results; therefore, by adjusting the parameters, the system performance can be further improved. Compared with the recognition rate mentioned above, the performance of the system has been greatly improved. This shows that T-S fuzzy neural network is effective in music recognition.

Figure 4

Recognition rate of piano music