To solve the problem existing in the measuring and prediction of transverse thickness distribution of finish hot rolling, the adaptive neural network trained by hybrid algorithms of particle swarm optimization (PSO) and back propagation (BP) neural network is introduced. Based on the BP network, the network structure, weights and threshold are optimized by PSO algorithm for improving the network convergence speed and prediction accuracy. By the data of two high reversible hot rolling mill, the average error of thickness is 3.6 μm and the error absolute value is less than 4 μm accounted for 87.1%. The absolute error frequency of strip thickness within 30 μm is 90% by statistical analysis of the steady state rolling, excepting the head and tail of the strip. The research results show that the network model can replace crown instrument to predict the transverse thickness distribution in the actual production. And the control means of strip shape are precisely controlled. It illustrates that the network model can meet the requirement of high precision flatness control.