tdnn(时间延迟神经网络).ppt

A time delay neural network algorithm for estimating image-pattern shape and motion Liu Xiao 2014.8.18 Image and Vision Computing, 1999 Christian Wohlera, Joachim K. Anlaufb institute of computing technology Liu Xiao 2014 8 18 Introduction In this paper, the author presents a novel concept for simultaneous shape estimation and motion analysis based on a TDNN(time-delay neural network) architecture with adaptable spatio-temporal receptive fields. Time delay network structure with adaptable spatio-temporal receptive fields t The input image sequence is of the dimension Sx(1) ╳ Sy(1) ╳ St(1) Time delay network structure with adaptable spatio-temporal receptive fields t t Bxyt rsmnp Branch Branch εsijt t Bxyt t Bxyt εsijt vsijqk σkt ωk Time delay network structure with adaptable spatio-temporal receptive fields Error function A k = c (current class c), A≈1 ?k = 0 k ≠ c Experiments The size of a single image is Sx(1) = 32 by Sy(1) = 16 pixels, one sequence consists of St(1) = 8 such images. There are two shape classes, and five speeds V0 = -4, -2, 0, 2, 4. The training sets consist of 5000 examples, 500 of each class. The network parameters are as follows: And the error rate on the test set is 1.1%. Experiments An Adaptable Time-Delay Neural-Network Algorithm for Image Sequence Analysis Liu Xiao 2014.8.18 IEEE TRANSACTIONS ON NEURAL NETWORKS, 1999 Introduction In this paper the author present an algorithm based on a time-delay neural network with spatio-temporal receptive fields and adaptable time delays(ATDNN) for image sequence analysis. The aim is to classify objects on temporal sequences of grayscale images and to estimate their motion behavior. The ATDNN algorithm The ATDNN algorithm The training algorithm As the adaptation procedure for ρ and β is based on a gradient descent method, next step is to define a network output ωk(ρ, β) for real-valued and , which is achieved by bilinear interpolation. Error functio