For the unstable output of the traditional texture classification meth ods based on extreme learning machine (ELM),this paper presents a new approach for the automatic classificat ion of texture images.In order to improve the generalization ability and the robustness of the learning model,thi s paper improves the traditional dynamical model by fusing the linear and nonlinear models together.Due to the fa ster learning speed,ELM is used as the basic classifier in this paper. Moreover,a proposed dynamical model is u tilized to realize the optimal fusion of multiple ELMs with the iteration of linear and local attractor.By combining multiple classifiers with the improved dynamical model,the corrupted classifier outputs are discarded accordin g to the classifier agreement.The experimental results on CUReT texture database demonstrate that the proposed app roach can improve the stability and classification accuracy significantly,and achieve a more ideal texture class ification compared with the traditional methods.