Scene recognition method, device and mobile terminal based on ambient sound

A method for mobile terminal locating according to antenna position locating and scene locating based on background sound information, comprising:

• collecting a background sound signal;
• converting the background sound signal in time domain into a frequency domain signal;
• calculating an arithmetic square for spectrum amplitude of the frequency domain signal to obtain power spectrum of the frequency domain signal, filtering the power spectrum of the frequency domain signal, calculating a logarithm for the filtered power spectrum of the frequency domain signal to obtain a result approximate to homomorphism conversion, and finally extracting a sound feature vector through a discrete cosine transform;
• inputting the sound feature vector under a neural network, obtaining a recognition result by calculating using intermediate matrixes obtained by using the neural network and the sound feature vector, and determining a scene corresponding to the sound feature vector;
• locating a mobile terminal by performing the scene locating according to the determined scene and the antenna position locating, to be complementary to the antenna position locating, to improve locating accuracy and improve performance of the mobile terminal;
• wherein, the neural network includes d input layer units, n mode layer units and c classification layer units, and
• the step of obtaining the intermediate matrixes comprises:
• normalizing samples x which are participated training to a unit length,
• ∑
• i
• =
• 1
• d
• ⁢
• x
• i
• 2
• =
• 1
• ,
• and placing the normalized samples on the input layer units;
• initializing connections connecting the input units and the mode layer units as wk=xk, wherein k=1, 2, 3, . . . , n;
• and establishing connections from the mode layer units to the corresponding classification layer units representing the classification to which the xk belongs, to form the neural network in which the input layer units and the mode layer units are completely connected and the mode layer units to the classification layer units are coefficient-connected;
• obtaining connection weight coefficients wjk which represent a component xjk connecting to the jth mode layer unit, wherein, xjk is the kth component of the jth sample, j=1, 2, . . . , n, k=1, 2, . . . , d, then obtaining a matrix W as the intermediate matrix.

A mobile terminal, comprising:

• a sound collection module, a preprocessing module, a feature extraction module, and a processing module;
• wherein,
• the sound collection module is configured to collect a background sound signal and transmit the background sound signal to the preprocessing module;
• the preprocessing module is configured to convert the background sound signal in time domain into a frequency domain signal and transmit the frequency domain signal to the feature extraction module;
• the feature extraction module is configured to:
• calculate an arithmetic square for spectrum amplitude of the frequency domain signal to obtain power spectrum of the frequency domain signal, filter the power spectrum of the frequency domain signal, calculate a logarithm for the filtered power spectrum of the frequency domain signal to obtain a result approximate to homomorphism conversion, and finally extract a sound feature vector through a discrete cosine transform;
• and
• the processing module is configured to receive the sound feature vector and obtain a scene corresponding to the sound feature vector and locate the mobile terminal by performing the scene locating according to the determined scene and the antenna position locating, to be complementary to the antenna position locating, to improve locating accuracy and improve performance of the mobile terminal;
• wherein the processing module comprises a scene recognition module and a database;
• wherein,
• the database is configured to store intermediate matrixes for different scenes which are obtained by using a neural network which includes d input layer units, n mode layer units and c classification layer units, wherein the intermediate matrixes are obtained by:
• normalizing samples x which are participated training to a unit length,
• ∑
• i
• =
• 1
• d
• ⁢
• x
• i
• 2
• =
• 1
• ,
• and placing the normalized samples on the input layer units;
• initializing connections connecting the input units and the mode layer units as wk=xk, wherein k=1, 2, 3, . . . , n;
• and establishing connections from the mode layer units to the corresponding classification layer units representing the classification to which the xk belongs, to form the neural network in which the input layer units and the mode layer units are completely connected and the mode layer units to the classification layer units are coefficient-connected;
• obtaining connection weight coefficients wjk which represent a component xjk connecting to the jth mode layer unit, wherein, xjk is the kth component of the jth sample, j=1, 2, . . . , n, k=1, 2, . . . , d, then obtaining a matrix W as the intermediate matrix;
• and
• the scene recognition module is configured to receive the sound feature vector, input the sound feature vector under the neural network, obtain a recognition result by calculating using the intermediate matrixes and the sound feature vector, and determine a scene corresponding to the sound feature vector.