The objective of the thesis is the development of a set of innovative algorithms around the topic of beamforming in the field of acoustic imaging, audio and image processing, aimed at significantly improving the performance of devices that exploit these computational approaches. Therefore the context is the improvement of devices (ultrasound machines and video/audio devices) already on the market or the development of new ones which, through the proposed studies, can be introduced on new the markets with the launch of innovative high-tech start-ups. This is the motivation and the leitmotiv behind the doctoral work carried out. In fact, in the first part of the work an innovative image reconstruction algorithm in the field of ultrasound biomedical imaging is presented, which is connected to the development of such equipment that exploits the computing opportunities currently offered nowadays at low cost by GPUs (Moore’s law). The proposed target is to obtain a new pipeline of the reconstruction of the image abandoning the architecture of such hardware based In the first part of the thesis I faced the topic of the reconstruction of ultrasound images for applications hypothesized on a software based device through image reconstruction algorithms processed in the frequency domain. An innovative beamforming algorithm based on seismic migration is presented, in which a transformation of the RF data is carried out and the reconstruction algorithm can evaluate a masking of the k-space of the data, speeding up the reconstruction process and reducing the computational burden. The analysis and development of the algorithms responsible for carrying out the thesis has been approached from a feasibility point in an off-line context and on the Matlab platform, processing both synthetic simulated generated data and real RF data: the subsequent development of these algorithms within of the future ultrasound biomedical equipment will exploit an high-performance computing framework capable of processing customized kernel pipelines (henceforth called ’filters’) on CPU/GPU. The type of filters implemented involved the topic of Plane Wave Imaging (PWI), an alternative method of acquiring the ultrasound image compared to the state of the art of the traditional standard B-mode which currently exploit sequential sequence of insonification of the sample under examination through focused beams transmitted by the probe channels. The PWI mode is interesting and opens up new scenarios compared to the usual signal acquisition and processing techniques, with the aim of making signal processing in general and image reconstruction in particular faster and more flexible, and increasing importantly the frame rate opens up and improves clinical applications. The innovative idea is to introduce in an offline seismic reconstruction algorithm for ultrasound imaging a further filter, named masking matrix. The masking matrices can be computed offline knowing the system parameters, since they do not depend from acquired data. Moreover, they can be pre-multiplied to propagation matrices, without affecting the overall computational load. Subsequently in the thesis, the topic of beamforming in audio processing on super-direct linear arrays of microphones is addressed. The aim is to make an in depth analysis of two main families of data-independent approaches and algorithms present in the literature by comparing their performances and the trade-off between directivity and frequency invariance, which is not yet known at to the state-of-the-art. The goal is to validate the best algorithm that allows, from the perspective of an implementation, to experimentally verify performance, correlating it with the characteristics and error statistics. Frequency-invariant beam patterns are often required by systems using an array of sensors to process broadband signals. In some experimental conditions, the array spatial aperture is shorter than the involved wavelengths. In these conditions, superdirective beamforming is essential for an efficient system. I present a comparison between two methods that deal with a data-independent beamformer based on a filter-and-sum structure. Both methods (the first one numerical, the second one analytic) formulate a mathematical convex minimization problem, in which the variables to be optimized are the filters coefficients or frequency responses. In the described simulations, I have chosen a geometry and a set-up of parameters that allows us to make a fair comparison between the performances of the two different design methods analyzed. In particular, I addressed a small linear array for audio capture with different purposes (hearing aids, audio surveillance system, video-conference system, multimedia device, etc.). The research activity carried out has been used for the launch of a high-tech device through an innovative start-up in the field of glasses/audio devices (https://acoesis.com/en/). It has been proven that the proposed algorithm gives the possibility of obtaining higher performances than the state of the art of similar algorithms, additionally providing the possibility of connecting directivity or better generalized directivity to the statistics of phase errors and gain of sensors, extremely important in superdirective arrays in the case of real and industrial implementation. Therefore, the method proposed by the comparison is innovative because it quantitatively links the physical construction characteristics of the array to measurable and experimentally verifiable quantities, making the real implementation process controllable. The third topic faced is the reconstruction of the Room Impluse Response (RIR) using audio processing blind methods. Given an unknown audio source, the estimation of time differences-of-arrivals (TDOAs) can be efficiently and robustly solved using blind channel identification and exploiting the cross-correlation identity (CCI). Prior blind works have improved the estimate of TDOAs by means of different algorithmic solutions and optimization strategies, while always sticking to the case N = 2 microphones. But what if we can obtain a direct improvement in performance by just increasing N? In the fourth Chapter I tried to investigate this direction, showing that, despite the arguable simplicity, this is capable of (sharply) improving upon state-of-the-art blind channel identification methods based on CCI, without modifying the computational pipeline. Inspired by our results, we seek to warm up the community and the practitioners by paving the way (with two concrete, yet preliminary, examples) towards joint approaches in which advances in the optimization are combined with an increased number of microphones, in order to achieve further improvements. Sound source localisation applications can be tackled by inferring the time-difference-of-arrivals (TDOAs) between a sound-emitting source and a set of microphones. Among the referred applications, one can surely list room-aware sound reproduction, room geometry’s estimation, speech enhancement. Despite a broad spectrum of prior works estimate TDOAs from a known audio source, even when the signal emitted from the acoustic source is unknown, TDOAs can be inferred by comparing the signals received at two (or more) spatially separated microphones, using the notion of cross-corrlation identity (CCI). This is the key theoretical tool, not only, to make the ordering of microphones irrelevant during the acquisition stage, but also to solve the problem as blind channel identification, robustly and reliably inferring TDOAs from an unknown audio source. However, when dealing with natural environments, such “mutual agreement” between microphones can be tampered by a variety of audio ambiguities such as ambient noise. Furthermore, each observed signal may contain multiple distorted or delayed replicas of the emitting source due to reflections or generic boundary effects related to the (closed) environment. Thus, robustly estimating TDOAs is surely a challenging problem and CCI-based approaches cast it as single-input/multi-output blind channel identification. Such methods promote robustness in the estimate from the methodological standpoint: using either energy-based regularization, sparsity or positivity constraints, while also pre-conditioning the solution space. Last but not least, the Acoustic Imaging is an imaging modality that exploits the propagation of acoustic waves in a medium to recover the spatial distribution and intensity of sound sources in a given region. Well known and widespread acoustic imaging applications are, for example, sonar and ultrasound. There are active and passive imaging devices: in the context of this thesis I consider a passive imaging system called Dual Cam that does not emit any sound but acquires it from the environment. In an acoustic image each pixel corresponds to the sound intensity of the source, the whose position is described by a particular pair of angles and, in the case in which the beamformer can, as in our case, work in near-field, from a distance on which the system is focused. In the last part of this work I propose the use of a new modality characterized by a richer information content, namely acoustic images, for the sake of audio-visual scene understanding. Each pixel in such images is characterized by a spectral signature, associated to a specific direction in space and obtained by processing the audio signals coming from an array of microphones. By coupling such array with a video camera, we obtain spatio-temporal alignment of acoustic images and video frames. This constitutes a powerful source of self-supervision, which can be exploited in the learning pipeline we are proposing, without resorting to expensive data annotations. However, since 2D planar arrays are cumbersome and not as widespread as ordinary microphones, we propose that the richer information content of acoustic images can be distilled, through a self-supervised learning scheme, into more powerful audio and visual feature representations. The learnt feature representations can then be employed for downstream tasks such as classification and cross-modal retrieval, without the need of a microphone array. To prove that, we introduce a novel multimodal dataset consisting in RGB videos, raw audio signals and acoustic images, aligned in space and synchronized in time. Experimental results demonstrate the validity of our hypothesis and the effectiveness of the proposed pipeline, also when tested for tasks and datasets different from those used for training. Chapter 6 closes the thesis, presenting a development activity of a new Dual Cam POC to build-up from it a spin-off, assuming to apply for an innovation project for hi-tech start- ups (such as a SME instrument H2020) for a 50Keuro grant, following the idea of the technology transfer. A deep analysis of the reference market, technologies and commercial competitors, business model and the FTO of intellectual property is then conducted. Finally, following the latest technological trends (https://www.flir.eu/products/si124/) a new version of the device (planar audio array) with reduced dimensions and improved technical characteristics is simulated, simpler and easier to use than the current one, opening up new interesting possibilities of development not only technical and scientific but also in terms of business fallout.
Advanced algorithms for audio and image processing
GRECO, DANILO
2021-02-25
Abstract
The objective of the thesis is the development of a set of innovative algorithms around the topic of beamforming in the field of acoustic imaging, audio and image processing, aimed at significantly improving the performance of devices that exploit these computational approaches. Therefore the context is the improvement of devices (ultrasound machines and video/audio devices) already on the market or the development of new ones which, through the proposed studies, can be introduced on new the markets with the launch of innovative high-tech start-ups. This is the motivation and the leitmotiv behind the doctoral work carried out. In fact, in the first part of the work an innovative image reconstruction algorithm in the field of ultrasound biomedical imaging is presented, which is connected to the development of such equipment that exploits the computing opportunities currently offered nowadays at low cost by GPUs (Moore’s law). The proposed target is to obtain a new pipeline of the reconstruction of the image abandoning the architecture of such hardware based In the first part of the thesis I faced the topic of the reconstruction of ultrasound images for applications hypothesized on a software based device through image reconstruction algorithms processed in the frequency domain. An innovative beamforming algorithm based on seismic migration is presented, in which a transformation of the RF data is carried out and the reconstruction algorithm can evaluate a masking of the k-space of the data, speeding up the reconstruction process and reducing the computational burden. The analysis and development of the algorithms responsible for carrying out the thesis has been approached from a feasibility point in an off-line context and on the Matlab platform, processing both synthetic simulated generated data and real RF data: the subsequent development of these algorithms within of the future ultrasound biomedical equipment will exploit an high-performance computing framework capable of processing customized kernel pipelines (henceforth called ’filters’) on CPU/GPU. The type of filters implemented involved the topic of Plane Wave Imaging (PWI), an alternative method of acquiring the ultrasound image compared to the state of the art of the traditional standard B-mode which currently exploit sequential sequence of insonification of the sample under examination through focused beams transmitted by the probe channels. The PWI mode is interesting and opens up new scenarios compared to the usual signal acquisition and processing techniques, with the aim of making signal processing in general and image reconstruction in particular faster and more flexible, and increasing importantly the frame rate opens up and improves clinical applications. The innovative idea is to introduce in an offline seismic reconstruction algorithm for ultrasound imaging a further filter, named masking matrix. The masking matrices can be computed offline knowing the system parameters, since they do not depend from acquired data. Moreover, they can be pre-multiplied to propagation matrices, without affecting the overall computational load. Subsequently in the thesis, the topic of beamforming in audio processing on super-direct linear arrays of microphones is addressed. The aim is to make an in depth analysis of two main families of data-independent approaches and algorithms present in the literature by comparing their performances and the trade-off between directivity and frequency invariance, which is not yet known at to the state-of-the-art. The goal is to validate the best algorithm that allows, from the perspective of an implementation, to experimentally verify performance, correlating it with the characteristics and error statistics. Frequency-invariant beam patterns are often required by systems using an array of sensors to process broadband signals. In some experimental conditions, the array spatial aperture is shorter than the involved wavelengths. In these conditions, superdirective beamforming is essential for an efficient system. I present a comparison between two methods that deal with a data-independent beamformer based on a filter-and-sum structure. Both methods (the first one numerical, the second one analytic) formulate a mathematical convex minimization problem, in which the variables to be optimized are the filters coefficients or frequency responses. In the described simulations, I have chosen a geometry and a set-up of parameters that allows us to make a fair comparison between the performances of the two different design methods analyzed. In particular, I addressed a small linear array for audio capture with different purposes (hearing aids, audio surveillance system, video-conference system, multimedia device, etc.). The research activity carried out has been used for the launch of a high-tech device through an innovative start-up in the field of glasses/audio devices (https://acoesis.com/en/). It has been proven that the proposed algorithm gives the possibility of obtaining higher performances than the state of the art of similar algorithms, additionally providing the possibility of connecting directivity or better generalized directivity to the statistics of phase errors and gain of sensors, extremely important in superdirective arrays in the case of real and industrial implementation. Therefore, the method proposed by the comparison is innovative because it quantitatively links the physical construction characteristics of the array to measurable and experimentally verifiable quantities, making the real implementation process controllable. The third topic faced is the reconstruction of the Room Impluse Response (RIR) using audio processing blind methods. Given an unknown audio source, the estimation of time differences-of-arrivals (TDOAs) can be efficiently and robustly solved using blind channel identification and exploiting the cross-correlation identity (CCI). Prior blind works have improved the estimate of TDOAs by means of different algorithmic solutions and optimization strategies, while always sticking to the case N = 2 microphones. But what if we can obtain a direct improvement in performance by just increasing N? In the fourth Chapter I tried to investigate this direction, showing that, despite the arguable simplicity, this is capable of (sharply) improving upon state-of-the-art blind channel identification methods based on CCI, without modifying the computational pipeline. Inspired by our results, we seek to warm up the community and the practitioners by paving the way (with two concrete, yet preliminary, examples) towards joint approaches in which advances in the optimization are combined with an increased number of microphones, in order to achieve further improvements. Sound source localisation applications can be tackled by inferring the time-difference-of-arrivals (TDOAs) between a sound-emitting source and a set of microphones. Among the referred applications, one can surely list room-aware sound reproduction, room geometry’s estimation, speech enhancement. Despite a broad spectrum of prior works estimate TDOAs from a known audio source, even when the signal emitted from the acoustic source is unknown, TDOAs can be inferred by comparing the signals received at two (or more) spatially separated microphones, using the notion of cross-corrlation identity (CCI). This is the key theoretical tool, not only, to make the ordering of microphones irrelevant during the acquisition stage, but also to solve the problem as blind channel identification, robustly and reliably inferring TDOAs from an unknown audio source. However, when dealing with natural environments, such “mutual agreement” between microphones can be tampered by a variety of audio ambiguities such as ambient noise. Furthermore, each observed signal may contain multiple distorted or delayed replicas of the emitting source due to reflections or generic boundary effects related to the (closed) environment. Thus, robustly estimating TDOAs is surely a challenging problem and CCI-based approaches cast it as single-input/multi-output blind channel identification. Such methods promote robustness in the estimate from the methodological standpoint: using either energy-based regularization, sparsity or positivity constraints, while also pre-conditioning the solution space. Last but not least, the Acoustic Imaging is an imaging modality that exploits the propagation of acoustic waves in a medium to recover the spatial distribution and intensity of sound sources in a given region. Well known and widespread acoustic imaging applications are, for example, sonar and ultrasound. There are active and passive imaging devices: in the context of this thesis I consider a passive imaging system called Dual Cam that does not emit any sound but acquires it from the environment. In an acoustic image each pixel corresponds to the sound intensity of the source, the whose position is described by a particular pair of angles and, in the case in which the beamformer can, as in our case, work in near-field, from a distance on which the system is focused. In the last part of this work I propose the use of a new modality characterized by a richer information content, namely acoustic images, for the sake of audio-visual scene understanding. Each pixel in such images is characterized by a spectral signature, associated to a specific direction in space and obtained by processing the audio signals coming from an array of microphones. By coupling such array with a video camera, we obtain spatio-temporal alignment of acoustic images and video frames. This constitutes a powerful source of self-supervision, which can be exploited in the learning pipeline we are proposing, without resorting to expensive data annotations. However, since 2D planar arrays are cumbersome and not as widespread as ordinary microphones, we propose that the richer information content of acoustic images can be distilled, through a self-supervised learning scheme, into more powerful audio and visual feature representations. The learnt feature representations can then be employed for downstream tasks such as classification and cross-modal retrieval, without the need of a microphone array. To prove that, we introduce a novel multimodal dataset consisting in RGB videos, raw audio signals and acoustic images, aligned in space and synchronized in time. Experimental results demonstrate the validity of our hypothesis and the effectiveness of the proposed pipeline, also when tested for tasks and datasets different from those used for training. Chapter 6 closes the thesis, presenting a development activity of a new Dual Cam POC to build-up from it a spin-off, assuming to apply for an innovation project for hi-tech start- ups (such as a SME instrument H2020) for a 50Keuro grant, following the idea of the technology transfer. A deep analysis of the reference market, technologies and commercial competitors, business model and the FTO of intellectual property is then conducted. Finally, following the latest technological trends (https://www.flir.eu/products/si124/) a new version of the device (planar audio array) with reduced dimensions and improved technical characteristics is simulated, simpler and easier to use than the current one, opening up new interesting possibilities of development not only technical and scientific but also in terms of business fallout.File | Dimensione | Formato | |
---|---|---|---|
phdunige_1620621_2.pdf
accesso aperto
Tipologia:
Tesi di dottorato
Dimensione
16.33 MB
Formato
Adobe PDF
|
16.33 MB | Adobe PDF | Visualizza/Apri |
phdunige_1620621_1.pdf
accesso aperto
Tipologia:
Tesi di dottorato
Dimensione
16.4 MB
Formato
Adobe PDF
|
16.4 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.