Image restoration algorithm for terahertz FMCW radar imaging.

The terahertz frequency modulation continuous-wave (THz FMCW) imaging technology has been widely used in non-destructive testing applications. However, THz FMCW real-aperture radar usually has a small depth of field and poor lateral resolution, thus restricting the high-precision imaging application. This paper proposes a 150-220 GHz FMCW Bessel beam imaging system, effectively doubling the depth of field and unifying the lateral resolution compared to the Gaussian beam quasi-optical system. Moreover, a THz image restoration algorithm based on local gradients and convolution kernel priors is proposed to eliminate further the convolution effect introduced by the Bessel beam, thereby enhancing the lateral resolution to 2 mm. It effectively improves the image under-restoration or over-restoration caused by the mismatch between the ideal and actual point spread function. The imaging results of the resolution test target and semiconductor device verify the advantages of the proposed system and algorithm.

High range resolution wideband terahertz FMCW radar with a large depth of field.

Terahertz frequency modulation continuous wave (THz FMCW) imaging technology has been widely used in non-destructive testing (NDT) applications of non-metallic materials. However, THz FMCW real-aperture radar usually has a narrow bandwidth and small depth of field, thus restricting the application of THz FMCW NDT. In this paper, a wideband THz signal (220-500 GHz) generation method is proposed by time-division multiplexing. Moreover, a dual-band quasi-optical design with a large depth of field is proposed based on the THz Bessel beam, and a high-quality range profile is obtained. Especially, a signal fusion extended Fourier analysis algorithm without prior knowledge is proposed to further enhance the range profile accuracy, which improves the range resolution to 0.28 mm (λ/3, center frequency 360 GHz). The effectiveness and advantages of the proposed system are verified by artificially constructing composite materials.

Balanced Thermal Insulation, Flame-Retardant and Mechanical Properties of PU Foam Constructed via Cost-Effective EG/APP/SA Ternary Synergistic Modification.

To address the challenge of balancing the mechanical, thermal insulation, and flame-retardant properties of building insulation materials, this study presented a facile approach to modify the rigid polyurethane foam composites (RPUFs) via commercial expandable graphite (EG), ammonium polyphosphate (APP), and silica aerogel (SA). The resulting EG/APP/SA/RPUFs exhibited low thermal conductivity close to neat RPUF. However, the compressive strength of the 6EG/2APP/SA/RPUF increased by 49% along with achieving a V-0 flame retardant rating. The residual weight at 700 °C increased from 19.2 wt.% to 30.9 wt.%. Results from cone calorimetry test (CCT) revealed a 9.2% reduction in total heat release (THR) and a 17.5% decrease in total smoke production (TSP). The synergistic flame-retardant mechanism of APP/EG made significant contribution to the excellent flame retardant properties of EG/APP/SA/RPUFs. The addition of SA played a vital role in reducing thermal conductivity and enhancing mechanical performance, effectively compensating for the shortcomings of APP/EG. The cost-effective EG/APP/SA system demonstrates a positive ternary synergistic effect in achieving a balance in RPUFs properties. This study provides a novel strategy aimed at developing affordable building wall insulation material with enhanced safety features.

ModeRNN: Harnessing Spatiotemporal Mode Collapse in Unsupervised Predictive Learning.

Learning predictive models for unlabeled spatiotemporal data is challenging in part because visual dynamics can be highly entangled, especially in real scenes. In this paper, we refer to the multi-modal output distribution of predictive learning as spatiotemporal modes. We find an experimental phenomenon named spatiotemporal mode collapse (STMC) on most existing video prediction models, that is, features collapse into invalid representation subspaces due to the ambiguous understanding of mixed physical processes. We propose to quantify STMC and explore its solution for the first time in the context of unsupervised predictive learning. To this end, we present ModeRNN, a decoupling-aggregation framework that has a strong inductive bias of discovering the compositional structures of spatiotemporal modes between recurrent states. We first leverage a set of dynamic slots with independent parameters to extract individual building components of spatiotemporal modes. We then perform a weighted fusion of slot features to adaptively aggregate them into a unified hidden representation for recurrent updates. Through a series of experiments, we show high correlation between STMC and the fuzzy prediction results of future video frames. Besides, ModeRNN is shown to better mitigate STMC and achieve the state of the art on five video prediction datasets.

VideoDG: Generalizing Temporal Relations in Videos to Novel Domains.

This paper introduces video domain generalization where most video classification networks degenerate due to the lack of exposure to the target domains of divergent distributions. We observe that the global temporal features are less generalizable, due to the temporal domain shift that videos from other unseen domains may have an unexpected absence or misalignment of the temporal relations. This finding has motivated us to solve video domain generalization by effectively learning the local-relation features of different timescales that are more generalizable, and exploiting them along with the global-relation features to maintain the discriminability. This paper presents the VideoDG framework with two technical contributions. The first is a new deep architecture named the Adversarial Pyramid Network, which improves the generalizability of video features by capturing the local-relation, global-relation, and cross-relation features progressively. On the basis of pyramid features, the second contribution is a new and robust approach of adversarial data augmentation that can bridge different video domains by improving the diversity and quality of augmented data. We construct three video domain generalization benchmarks in which domains are divided according to different datasets, different consequences of actions, or different camera views, respectively. VideoDG consistently outperforms the combinations of previous video classification models and existing domain generalization methods on all benchmarks.