Binaural sound source localization (BSSL) in low signal-to-noise ratio (SNR) and high reverberation environment is still a challenging task. In this paper, a novel BSSL algorithm is proposed by introducing convolutional neural network (CNN). The proposed algorithm first extracts the spatial feature of each sub-band from binaural sound signal, and then combines the features of all sub-bands within one frame to assemble a two-dimensional feature matrix as a grey image. To fully exploit the advantage of the CNN in extracting high-level features from the grey image, the spatial feature matrix of each frame is used as input to train the CNN model. The CNN is then used to predict azimuth of sound source. The experiments show that the proposed algorithm significantly improves the localization performance of BSSL in various acoustic environments, especially to deal with low SNR and high reverberation conditions.
In this article, a tree-indexed demarcation using forward and backward searches is proposed to segment the seriously touching characters, which is used as an important procedure for the character recognition in the real industrial scenarios. Unlike the segmentation directly depending on the histogram analysis, the reliable and unreliable valleys are first identified from a smoothed projection curve in our method, where a set of image pre-procedures are carefully carried out to catch this smoothed curve. With these operations, two segment binary trees are sequently constructed via the corresponding forwarding and backward searches of the valleys. Then, the delimiter candidates are sophisticatedly selected with a number of well-designed rules. The optimal segmented paths are finally obtained on the width variation measure of the characters, and the isolated characters are achieved. Experiments show that with our tree-indexed demarcation using forward and backward searches method, the touching character groups can be effectively segmented to better serve the industrial automatic recognition system.
Heart sounds are highly valuable to cardiovascular diseases in clinical diagnoses. So the analysis of Phonocardiogram (PCG) is helpful in the field. It is significant to position heart sounds in time domain accurately and automatically. A time-frequency transformation based on Matching Pursuit Method (MPM) is developed for the PCG. The method was applied to 60 cardiac cycles. It positioned the first (S1) and the second (S2) heart sounds automatically. Also their averages were calculated. The results show that MPM gives a correct detection rate up to 93.3%. The processing is real-time. The method can eliminate Electrocardiogram (ECG) as time-reference signal and reduce hardware overheads. It is independent of subjective human judgment and robust to artifacts, background noise and other interferences.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.