Future video frame prediction based on generative motion-assistant discriminative network

The drastic variation of motion in spatial and temporal dimensions makes the video prediction task extremely challenging. Existing RNN models obtain higher performance by deepening or widening the model. They obtain the multi-scale features of the video only by stacking layers, which is inefficient and brings unbearable training costs (such as memory, FLOPs, and training time). Different from them, this paper proposes a spatiotemporal multi-scale model called MS-LSTM wholly from a multi-scale perspective. On the basis of stacked layers, MS-LSTM incorporates two additional efficient multi-scale designs to fully capture spatiotemporal context information. Concretely, we employ LSTMs with mirrored pyramid structures to construct spatial multi-scale representations and LSTMs with different convolution kernels to construct temporal multi-scale representations. Detailed comparison experiments with eight baseline models on four video datasets show that MS-LSTM has better performance but lower training costs.

MS-LSTM: Exploring Spatiotemporal Multiscale Representations in Video Prediction Domain

Video frame prediction with dual-stream deep network emphasizing motions and content details

One crucial problem with applying digital image analysis technology to the field of agriculture is separation of green vegetation regions. In this paper, a new crop canopy image segmentation algorithm, which is based on Fisher linear discriminant and the values of red, green and blue of each pixel in an image, is presented. First, 90 green vegetation regions and 80 non-green vegetation regions, each of which has nine pixels, from two crop canopy images were chosen to generate training data. Then, the optimal projection direction is determined by using the Fisher linear discriminant. Finally, a color crop canopy image is divided into two parts—green vegetation and non-green vegetation---with a fixed threshold. The algorithm’s performance was assessed on 50 images. The results on 50 images show that the median of mis-segmentation of proposed method is about 5%. The comparisons between the proposed algorithm and those based on color indices show that the former outperforms the latter with high segmentation rate and fast running speed.

Segmentation of Green Vegetation from Crop Canopy Images Based on Fisher Linear Discriminant

Knowledge of the microphysical characteristics of precipitation plays a significant role in meteorology, hydrology, and natural hazards management, especially in the western Sichuan Basin (WSB), which is located east of the Tibetan Plateau (TP) in southwestern China and thus has unique terrain conditions and weather systems. Nonetheless, the literature regarding raindrop size distribution (RSD) in the WSB is still very limited. This work investigates RSD characteristics and temporal variations in a site (Chengdu, CD) of the WSB by employing three years of quality-controlled RSD observation collected from a second-generation PARSIVEL disdrometer. The results show that RSD has noticeable seasonal and diurnal variations in CD. Specifically, the broadest mean raindrop spectra can be found in summer and the narrowest in winter, and the raindrop spectra of a day can be the narrowest during 1400–1500 BJT (Beijing Standard Time, UTC+8). In addition, the mass-weighted mean diameter (Dm) is lower in the daytime than in the nighttime, while the logarithm of the generalized intercept parameter (log10Nw, the unit of the Nw is m−3 mm−1) has a larger value in the daytime than in the nighttime. In addition, intercomparisons indicate that the mean Dm of convective rains in CD is smaller than in South China and it is higher than in the eastern slope of TP, East China, and North China; on the other hand, the corresponding mean log10Nw is close to the value at the middle TP. Local empirical relations of shape–slope parameters (μ–Λ) and reflectivity–rain rate (Z–R) are also presented to provide references for optimizing the RSD parameterization scheme and radar precipitation estimation in the local area.

/pdf/characteristics-and-variations-of-raindrop-size-distribution-2oy38q29.pdf

Li Ying Zheng

Papers

Video frame prediction with dual-stream deep network emphasizing motions and content details

Segmentation of Green Vegetation from Crop Canopy Images Based on Fisher Linear Discriminant

Characteristics and Variations of Raindrop Size Distribution in Chengdu of the Western Sichuan Basin, China