Efficient deep neural network architectures with multi-receptive feature optimization for multi-dimensional data processing

Abstract

In this thesis, efficient deep neural network architectures are proposed for extracting features from diverse receptive fields by introducing numerous optimization, feature fusion, and data transformation schemes targeting numerous multi-dimensional ap- plications. Firstly, to exploit effective features in data constrained environment from a single modality, feature learning from multiple perspectives is introduced through varying resolutions and novel data augmentation strategies. In addition, features from varying receptive fields have been extracted by introducing multi- kernel depthwise separable convolutions with varying dilation rates, and the per- formance is validated in the case of 1-D (electrocardiogram, ECG) and 2-D (chest X-ray image) data for disease classification. Afterward, feature spaces of multiple modalities have been explored by incorporating various transformed representations of multi-modal 1D time series sensor data. Moreover, a sequential training algo- rithm is proposed to gradually converge extracted features to the final objective and the overall scheme is deployed on human activity recognition application. For the purpose of multi-dimensional data segmentation (2D endoscopy images, and 3D CT volumes), instead of using conventional uni-scale feature propagation, multi-scale contextual feature aggregation and fusion-based building blocks are designed and incorporated in the DNN which offers improved feature sharing while minimizing the contextual information loss. Especially in the case of 3D data, a hybrid DNN architecture is proposed performing 2D slice-wise processing accompanied by lighter 3D-volumetric segmentation to reduce the complexity of the optimization process. Finally, a triple attention based learning scheme is proposed combining the channel, spatial, and pixel level attentions, which is incorporated in the DNN architectures to improve the feature sharing process targeting multiple objectives through hierarchi- cal training and joint optimization. The proposed methods are validated in various multi-dimensional datasets targeting real-world applications such as disease clas- sification, infection segmentation, disease severity prediction, and human activity recognition.