Method for improving talkback tone quality of monitoring camera

Abstract

The invention belongs to the technical field of monitoring cameras, and particularly relates to a method for improving talkback tone quality of a monitoring camera. The method comprises the following steps of: firstly, widely collecting an audio data set containing pure voice and various environmental noises, and preprocessing and normalizing by using a frequency spectrum subtraction algorithm and a Wiener filtering algorithm; then, selecting a convolutional neural network (CNN) architecture to perform cross validation training on preprocessed data to optimize model parameters; in the key Wiener filtering link, the received audio is firstly subjected to direct current removal and trend term processing, the power spectral density is independently estimated in different periods or frequency bands, adaptive filtering dynamic adjustment and smoothing technology noise reduction are carried out, and a filter response is set according to an optimization ratio. Through series processing, noise is accurately filtered out, voice characteristics are preserved, voice definition, identifiability and real-time performance are greatly improved, talkback tone quality is remarkably optimized, and user experience and commanding and dispatching efficiency are powerfully improved.

Description

Technical Field

[0001] The invention belongs to the technical field of monitoring cameras, and in particular relates to a method for improving the intercom sound quality of monitoring cameras. Background Art

[0002] In the coal mine production monitoring scenario, the video monitoring system plays a key role. It can not only capture the production images of each working face in real time, but the voice intercom function is also an essential element for achieving efficient command and dispatch. However, the special network environment of coal mines has caused serious damage to the sound quality of the surveillance camera's voice intercom. On the one hand, the coal mine network architecture is complex, involving many network devices and transmission links. The signal is easily interfered during the transmission process, causing the voice signal to attenuate and distort. The electromagnetic interference generated by the operation of a large number of electrical equipment frequently changes the signal transmission characteristics, causing the voice waveform to distort. On the other hand, the hardware equipment covered by command and dispatch is diverse and the performance is uneven. Acoustic components such as microphones and speakers are eroded by the harsh environment underground, reducing sensitivity and changing frequency response characteristics, introducing additional noise and distortion.

[0003] In the existing technology, although there are denoising algorithms to deal with such problems, they have significant defects. In complex noise environments, the ability to identify and separate noise is insufficient, making it difficult to completely remove noise components, and residual noise interferes with voice clarity. At the same time, there is a significant delay in the processing process, which destroys the real-time and continuity of the voice, degrades the echo effect, and seriously affects the user's intercom experience and command and dispatch efficiency. In view of this, innovative methods are urgently needed to improve the sound quality of surveillance camera intercoms to meet the urgent needs of coal mine safety production for accurate and efficient voice communication. Summary of the invention

[0004] The present invention aims at the technical problems existing in the above-mentioned background technology and proposes a method for improving the sound quality of intercom of a surveillance camera.

[0005] In order to achieve the above object, the technical solution adopted by the present invention comprises the following steps:

[0006] S1. Collect audio datasets containing pure speech signals and various types of environmental noise;

[0007] S2, select convolutional neural network CNN as the model architecture;

[0008] S3. Input the audio signal to be processed into the trained CNN model. In the Wiener filtering process, the calculation of the power spectrum density is improved. The power spectrum density value of n groups of data is calculated and the calculation formula is as follows: Among them, S (f)is the representation of the estimated signal in the frequency domain, H(f) is the frequency response of the channel, Ps(f) is the power spectral density of the original signal in the frequency domain, Pn(f) is the power spectral density of the noise in the frequency domain, X(f) is the representation of the received signal in the frequency domain, T is the transpose, and W is the filter order.

[0009] Preferably, in step S1, the audio data is subjected to denoising and filtering preprocessing operations using a denoising algorithm, and the data is normalized so that the eigenvalues ​​of different data have the same scale.

[0010] Preferably, the processing of the CNN model architecture in step S1 is to use the preprocessed audio data set to train the CNN model, and the model performance is evaluated by cross-validation during the training process, and the model parameters and optimization algorithm are adjusted according to the evaluation results.

[0011] Preferably, the noise reduction algorithm includes spectrum subtraction and Wiener filtering, and distinguishes speech and noise components by analyzing the spectrum characteristics of the input signal.

[0012] Preferably, in step S3, before estimating the power spectral density, the data is preprocessed, including removing DC components and trend items, dividing the received signal into multiple shorter time periods or frequency segments, independently estimating the power spectral density in each segment, and dynamically adjusting the power spectral density estimate according to the currently received signal using adaptive filtering technology, using smoothing technology to reduce fluctuations and noise in the calculated signal and noise power spectral density ratio, and designing the frequency response H(f) of the filter based on the optimized power spectral density ratio.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are that it innovates the calculation of the power spectrum density of the Wiener filter, first pre-processes the data, removes the DC component and trend term, and independently estimates it in different time periods or frequency bands, and combines adaptive filtering dynamic adjustment and smoothing technology to reduce noise, thereby optimizing the filter frequency response. This effectively overcomes the limitations of traditional algorithms in complex environments, accurately separates speech and noise, and avoids residual noise interference. It greatly improves speech clarity and recognizability, enhances real-time and continuity, and eliminates processing delays. In noisy scenes such as coal mines, the sound quality of surveillance camera intercoms is effectively optimized, significantly improving user experience and command and dispatch efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0015] Figure 1 Optimize the flow chart for the algorithm; Figure 2 The spectrum diagram before and after denoising by the Wiener filtering algorithm. DETAILED DESCRIPTION

[0016] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0017] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways than those described herein. Therefore, the present invention is not limited to the specific embodiments of the following disclosure.

[0018] Embodiment, in order to process the sound quality of voice intercom in the video surveillance system, eliminate the problems of intercom sound distortion and high noise, so as to improve the user experience. The existing denoising algorithm will have problems of unclean noise removal and delay, which affects the echo effect. The present invention proposes a method for improving the sound quality of voice intercom in surveillance cameras. Through deep learning and artificial intelligence technology, the algorithm can more accurately identify and suppress environmental noise while retaining the integrity of the voice signal. The algorithm can adapt to different types of environmental noise and voice signals, and has better generalization and application capabilities. The optimized noise suppression algorithm can significantly improve the user experience, make the voice signal clearer and more recognizable, so as to achieve the best transmission efficiency, improve the sound quality of voice intercom in surveillance cameras, and greatly improve the user experience. Figure 1 Optimize the flow chart for the algorithm.

[0019] First, we collect audio data sets containing pure speech signals and various types of environmental noise (such as traffic noise, wind noise, human voice, etc.), and use noise reduction algorithms such as spectrum subtraction and Wiener filtering on the audio data. By analyzing the spectral characteristics of the input signal, we distinguish the speech and noise components, and perform preprocessing operations such as denoising and filtering to improve the data quality. At the same time, we normalize the data so that the eigenvalues ​​between different data have the same scale.

[0020] Convolutional Neural Network (CNN) was selected as the model architecture because CNN has good performance in audio signal processing. The CNN model was trained using the preprocessed audio dataset. During the training process, cross-validation was used to evaluate the performance of the model, and the model parameters and optimization algorithm were adjusted based on the evaluation results. Cross-validation is a technique for evaluating the performance of machine learning models. It divides the dataset into several small parts, and then uses one part as the test set and the rest as the training set in turn. In this way, we can obtain the performance of the model on different datasets. Cross-validation can simulate various situations that the model may encounter in actual applications, thereby helping us discover and correct problems that may exist in the model in specific situations.

[0021] The audio signal to be processed is input into the trained CNN model, and the model intelligently identifies and suppresses the noise component based on the characteristics of the input signal. Then the clean speech signal after noise suppression is output. The algorithm is further optimized for specific application scenarios (such as speech recognition). The model parameters can be adjusted to improve the clarity and recognizability of the speech signal. The model parameters, such as learning rate and batch size, can be optimized through grid search, random search and other methods. The algorithm calculates the signal-to-noise ratio, which is the power ratio of the speech signal to the noise signal. The calculation of the signal-to-noise ratio helps determine the degree of noise suppression.

[0022] As a classic optimal filtering technique, the core of Wiener filtering is to estimate the original signal by minimizing the mean square error of the signal. The performance of this process is highly dependent on the accurate estimation of the signal power spectral density Ps(f) and the noise power spectral density Pn(f). In practical applications, the estimation of these power spectral densities often faces various challenges, such as limited observation data, non-stationarity of noise, and complex interactions between signals and noise. Therefore, improving the estimation method of Ps(f) and Pn(f) is the key to improving the performance of Wiener filtering. If the estimation is inaccurate, it may lead to unsatisfactory filtering effect and may even introduce additional distortion. First, before estimating the power spectral density, the data is preprocessed, such as removing the DC component and the trend term, to reduce the estimation error. Then the received signal X(f) is divided into multiple shorter time periods or frequency segments, and the power spectral density is estimated independently in each segment, which helps to deal with non-stationary noise and signal changes and improve the accuracy of the estimation. Adaptive filtering techniques, such as recursive least squares (RLS) or Kalman filtering, are used to dynamically adjust the power spectral density estimate based on the current received signal. This method can better adapt to changes in signals and noise. When calculating the ratio of the signal and noise power spectral density at each frequency point, smoothing techniques (such as moving average or low-pass filtering) are used to reduce fluctuations and noise in the ratio. Finally, based on the optimized power spectral density ratio, the frequency response H(f) of the filter is designed. This usually involves converting the ratio into a gain or attenuation factor of the filter to ensure that signal information is retained at frequency points with high signal power and noise is suppressed at frequency points with high noise power. The performance of the improved Wiener filter is evaluated through simulation or actual testing. The signal quality before and after filtering (such as signal-to-noise ratio, mean square error, etc.) is compared and the frequency response of the filter is observed to see if it meets expectations. Based on the evaluation results, the power spectral density estimation method, filter design and implementation are iteratively optimized to further improve the performance of the Wiener filter. Therefore, the calculation of Ps(f) and Pn(f) is improved, and is calculated from the power spectral density values ​​of n groups of data. The calculation formula is as follows: Among them, S (f) is the representation of the estimated signal in the frequency domain, H(f) is the frequency response of the channel, Ps(f) is the power spectral density of the original signal in the frequency domain, Pn(f) is the power spectral density of the noise in the frequency domain, X(f) is the representation of the received signal in the frequency domain, T is the transpose, and W is the filter order. Figure 2 Spectrum diagram before and after denoising using the Wiener filter algorithm. Figure 2It can be seen that after the algorithm is improved by the present invention, the noise in the low-frequency and high-frequency parts of the recording is effectively filtered out. Through deep learning and artificial intelligence technology, the algorithm can more accurately identify and suppress environmental noise while retaining the integrity of the voice signal. The algorithm can adapt to different types of environmental noise and voice signals and has better generalization and application capabilities. The optimized noise suppression algorithm can significantly improve the user experience and make the voice signal clearer and more recognizable.

[0023] The above description is only a preferred embodiment of the present invention and does not limit the present invention in other forms. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A method for improving the sound quality of surveillance camera intercom, characterized in that: The following steps are involved: S1. Collect audio datasets containing pure speech signals and various types of environmental noise; S2, select convolutional neural network CNN as the model architecture; S3. Input the audio signal to be processed into the trained CNN model. In the Wiener filtering process, the calculation of the power spectrum density is improved. The power spectrum density value of n groups of data is calculated and the calculation formula is as follows: Among them, S (f) is the representation of the estimated signal in the frequency domain, H(f) is the frequency response of the channel, Ps(f) is the power spectral density of the original signal in the frequency domain, Pn(f) is the power spectral density of the noise in the frequency domain, X(f) is the representation of the received signal in the frequency domain, T is the transpose, and W is the filter order.

2. A method for improving the sound quality of surveillance camera intercom according to claim 1, characterized in that: In step S1, the audio data is subjected to denoising and filtering preprocessing operations using a denoising algorithm, and the data is normalized so that the eigenvalues ​​of different data have the same scale.

3. The method for improving the sound quality of surveillance camera intercom according to claim 1, characterized in that: The processing of the CNN model architecture in step S1 is to use the preprocessed audio data set to train the CNN model, and the model performance is evaluated by cross-validation during the training process, and the model parameters and optimization algorithm are adjusted according to the evaluation results.

4. The method for improving the sound quality of surveillance camera intercom according to claim 2, characterized in that: The noise reduction algorithm includes spectrum subtraction and Wiener filtering, and distinguishes speech and noise components by analyzing the spectrum characteristics of the input signal.

5. The method for improving the sound quality of surveillance camera intercom according to claim 1, characterized in that: In the step S3, before estimating the power spectral density, the data is preprocessed, including removing the DC component and the trend item, and the received signal is divided into multiple shorter time periods or frequency segments. The power spectral density is independently estimated in each segment, and the power spectral density estimation is dynamically adjusted according to the currently received signal using adaptive filtering technology, and the smoothing technology is used to reduce the fluctuation and noise in the calculated signal and noise power spectral density ratio. The frequency response H(f) of the filter is designed according to the optimized power spectral density ratio.

Images