Wake word (WW) spotting is challenging in far-field due to the complexities and variations in acoustic conditions and the environmental interference in signal transmission. A suite of carefully designed and optimized audio front-end (AFE) algorithms help mitigate these challenges and provide better quality audio signals to the downstream modules such as WW spotter. Since the WW model is trained with the AFE-processed audio data, its performance is sensitive to AFE variations, such as gain changes. In addition, when deploying to new devices, the WW performance is not guaranteed because the AFE is unknown to the WW model. To address these issues, we propose a novel approach to use a new feature called $\Delta$LFBE to decouple the AFE gain variations from the WW model. We modified the neural network architectures to accommodate the delta computation, with the feature extraction module unchanged. We evaluate our WW models using data collected from real household settings and showed the models with the $\Delta$LFBE is robust to AFE gain changes. Specifically, when AFE gain changes up to $\pm$12dB, the baseline CNN model lost up to relative 19.0% in false alarm rate or 34.3% in false reject rate, while the model with $\Delta$LFBE demonstrates no performance loss.

The diagnosis, comparative and population study of cardiac radiology data require heart segmentation on increasingly large amount of images from different modalities/chambers/patients under various imaging views. Most existing automatic cardiac segmentation methods are often limited to single image segmentation with regulated modality/region settings or well-cropped ROI areas, which is impossible for large datasets due to enormous device protocols and institutional differences. A pure data-driven unsupervised segmentation without regulated setting requirements is crucial in this scenario, and will significantly automate the manual work and adopt the various changes of modality, subject or view. In this paper, we propose a general unsupervised groupwise segmentation: a direct simultaneous segmentation for a group of multi-modality, multi-chamber, multi-subject ( M3) cardiac images from a freely chosen imaging view. The segmentation can directly perform not only on regulated two/four-chamber images, but also on non-regulated uncropped raw MR/CT scans. A new Synchronized Spectral Network (SSN) is developed for the simultaneous decomposing, synchronizing, and clustering the spectral features of free-view M3cardiac images. The SSN-based groupwise analysis of image spectral bases immediately leads to groupwise segmentation of M3freeview images. The segmentation is quantitatively evaluated by three datasets (MR and CT mixed) with more than 200 subjects. High dice metric ( ) is consistently achieved in validation. Our method provides a powerful tool for medical images under general imaging environment.

Computer-aided diagnosis of spine problems relies on the automatic identification of spine structures in images. The task of automatic vertebra recognition is to identify the global spine and local vertebra structural information such as spine shape, vertebra location and pose. Vertebra recognition is challenging due to the large appearance variations in different image modalities/views and the high geometric distortions in spine shape. Existing vertebra recognitions are usually simplified as vertebrae detections, which mainly focuses on the identification of vertebra locations and labels but cannot support further spine quantitative assessment. In this paper, we propose a vertebra recognition method using 3D deformable hierarchical model (DHM) to achieve cross-modality local vertebra location+pose identification with accurate vertebra labeling, and global 3D spine shape recovery. We recast vertebra recognition as deformable model matching, fitting the input spine images with the 3D DHM via deformations. The 3D model-matching mechanism provides a more comprehensive vertebra location+pose+label simultaneous identification than traditional vertebra location+label detection, and also provides an articulated 3D mesh model for the input spine section. Moreover, DHM can conduct versatile recognition on volume and multi-slice data, even on single slice. Experiments show our method can successfully extract vertebra locations, labels, and poses from multi-slice T1/T2 MR and volume CT, and can reconstruct 3D spine model on different image views such as lumbar, cervical, even whole spine. The resulting vertebra information and the recovered shape can be used for quantitative diagnosis of spine problems and can be easily digitalized and integrated in modern medical PACS systems.

The efficient and robust extraction of invariant patterns from an image is a long-standing problem in computer vision. Invariant structures are often related to repetitive or near-repetitive patterns. The perception of repetitive patterns in an image is strongly linked to the visual interpretation and composition of textures. Repetitive patterns are products of both repetitive structures as well as repetitive reflections or color patterns. In other words, patterns that exhibit near-stationary behavior provide rich information about objects, their shapes, and their texture in an image. In this paper, we propose a new algorithm for repetitive pattern detection and grouping. The algorithm follows the classical region growing image segmentation scheme. It utilizes a mean-shift-like dynamic to group local image patches into clusters. It exploits a continuous joint alignment to: 1) match similar patches, and 2) refine the subspace grouping. We also propose an algorithm for inferring the composition structure of the repetitive patterns. The inference algorithm constructs a data-driven structural completion field, which merges the detected repetitive patterns into specific global geometric structures. The result of higher level grouping for image patterns can be used to infer the geometry of objects and estimate the general layout of a crowded scene.

The efficient and robust extraction of invariant patterns from an image is a long-standing problem in computer vision. Invariant structures are often related to repetitive or near-repetitive patterns. The perception of repetitive patterns in an image is strongly linked to the visual interpretation and composition of textures. Repetitive patterns are products of both repetitive structures as well as repetitive reflections or color patterns. In other words, patterns that exhibit near-stationary behavior provide a rich information about objects, their shapes, and their texture in an image. In this paper, we propose a new algorithm for repetitive pattern detection and grouping. The algorithm follows the classical region growing image segmentation scheme. It utilizes a mean-shift-like dynamics to group local image patches into clusters. It exploits a continuous joint alignment to (a) match similar patches and (b) refine the subspace grouping. The result of higher-level grouping for image patterns can be used to infer the geometry of object surfaces and estimate the general layout of a crowded scene.

Visual patterns generated by color patches, texture regions, and repetitive textons in an image can be organized into higher-level structural forms such as geometric shapes, arrays, and partition groups. Understanding the information content formed by these visual pattern compositions is important both from a theoretical point of view as well as in the robust implementation of many image processing applications. In this paper we propose a new method for building pattern compositions and inferring the high-level structural forms over near regular textures. We exploit the shape geometry of repeated patterns to interpret pairwise connections between patterns and generate the abstract structural form by unifying the local connections. The inferred structure can reflect the organization of multiple repeated patterns and can be used in the classification of texture structures.