Search

Recently, deep neural networks need to be deployed with low memory and computing resources, so it is necessary to design an efficient and compact network structure. This paper proposed a model compression method (KE ) based on improved attention transfer for the des ign of compact neural networks, which mainly used a wide residual teacher network (WRN) to guide a compact student network (KENet) by extracting both spatial and channel-wise attention to improve the performance, and applied this method to real-time object detection. The image classification experiment on CIFAR verifies that the knowledge distillation method with improved attention transfer can improve the performance of the compact model. The object detection experiment on VOC verifies that the model KEDet has good accuracy (72. 7 mAP) and time performance (86 fps) . The experimental results show that the object detection model based on improved attention transfer has good accuracy and real-time performance. [ABSTRACT FROM AUTHOR]

Front vehicle ranging plays a vital role in the field of autonomous vehicle technology. Aiming at the problem that the current vehicle ranging technology based on monocular vision neglects the lower edge of the vehicle connected to the ground, this paper proposes a monocular vision ranging model based on vehicle lower edge estimation and inverse perspective transformation, which realizes highprecision distance measurement of lateral and longitudinal directions of vehicles in front. Firstly, the vehicle's key point estimation and geometric relationship model are used to estimate the lower edge of the vehicle, then the distance measurement key points are calculated, and the point-based inverse perspective transformation distance measurement model is used to calculate the distance. The experimental results show that, compared with other monocular vision vehicle ranging algorithms, this method improves the accuracy and stability of ranging. [ABSTRACT FROM AUTHOR]

In order to solve the problem that the speed and accuracy of the current similarity video retrieval methods still cannot meet the requirements of users, this paper proposed a new similarity video quick retrieval method, which combined the three-dimen sional convolutional neural network with the hash learning method and applied to video data. It could not only quickly learn the video spatiotemporal feature representation, but also greatly shorten the video retrieval time. The experimental results on the set show that the similarity retrieval performance of the video using the proposed method is superior to the current mainstream methods. [ABSTRACT FROM AUTHOR]

Aiming at the convergence performance of FastICA is easy to affected by initial value selection of the initial demixing matrix, this paper introduced the gradient descent method to reduce initial value sensitivity, and put forward the improved secant method to accelerate the convergence speed. The experimental results show that the improved algorithm compared with other FastICA algorithm, not only improves the separation performance, but also reduces the number of iterations and enhances the convergence stability. Therefore, the improved algorithm overcomes the sensitive influence of initial value selection, and achieves faster and more robust speech separation performance. [ABSTRACT FROM AUTHOR]

Aiming at the problems of inaccurate large target segmentation and loss of small target information in current real-time semantic segmentation methods, this paper proposed a real-time semantic segmentation algorithm based on multi-branch networks. First of all, this paper optimized the bilateral segmentation network, and designed pyramid branches to expand the receptive field to cover large objects in the field of view and fully combine context information; secondly, designed a bilateral guidance fusion module to map deep and shallow features Provide guidance information to make up for the loss of small target information. Finally, this paper verified the proposed method on the Cityscapes dataset. The experimental results show that the proposed model achieves an average intersection ratio of 77.8% at an inference speed of 51.3 FPS, and the accuracy is increased by 2.5 percentage points compared with the baseline. The proposed method adopts the pyramid branch to obtain the characteristics of semantic edge regions at different scales while expanding the receptive field, and enhance the modeling ability of semantic boundaries, and the proposed bilateral guidance fusion module can more effectively integrate features of different levels, Compensating for the information loss caused by downsampling can better guide model learning. [ABSTRACT FROM AUTHOR]

Most of the existing deep learning-based methods for hand pose estimation use a standard three -dimension convolutional neural network(3D CNN ) to extract 3D features and estimate the 3D coordinates of hand joints. The features extracted by these methods lack the multi-scale information of the hand, which limits the accuracy of hand pose estimation. In addition, due to the huge computational cost and memory requirements of the 3D CNN, these methods are often difficult to meet the realtime requirement. To overcome these weaknesses, the proposed method used a spatial filter and a depth filter to simulate 3D convolutions, which reduced the amount of parameters. It extracted and integrates features at various scales, making full use of the 3D information of hand pose. Experiments show that this method can improve estimation accuracy, reduce model size, and run at over 119 fps on a standard computer with a single GPU. [ABSTRACT FROM AUTHOR]

Multimodal fusion algorithm is one of the solutions to the problem of urban road video caption which only considers visual information and ignores the equally important audio information. Existing multimodal fusion algorithms based on Transformer all have the problem of low fusion performance between modes and high computational complexity. In order to improve the interaction between multimodal information, This paper recently proposed a new Transformer based model called Multimodal Attention Bottleneck for Video Captioning (MABVC). Firstly, this paper uses pre-trained I3 D and VGGish networks to extract visual and audio features of video and input the extracted features into Transformer model. Then, the decoder part will train the information of the two modes respectively and perform multimodal fusion. Finally, the model processes the results of the decoder and generates text captions that people can understand. This paper conducted a comparison experiments using data sets MSR-VTT, MSVD and self-built data sets BUUISE, and validated model results using evaluation metrics the model. The experimental results show that the video caption model based on multimodal attention fusion has obvious improvement in all indicators. The model can still achieve good results on traffic scene data sets, and has great application prospects/can be promoted and applied in intelligent driving industry. [ABSTRACT FROM AUTHOR]