To eliminate the local artifacts in terahertz (THz) images caused by power fluctuation effect, a THz image enhancement model based on homomorphic filtering is constructed. However, the parameter values of the enhancement model have large differences and strong coupling, which brings great difficulties to determine the parameters of the enhancement model. Therefore, an adaptive teachinglearning-based optimization algorithm based on chaotic mapping and differential evolution is proposed to solve the optimal parameters of the enhancement model. Firstly, the standard Logistic chaotic mapping is improved, which increases the population diversity. Secondly, the update rate of fitness is introduced, the adaptive adjustment function of the inertial weight is constructed and the global and local optimization abilities are balanced, which is beneficial for the population to approach the optimal solution Thirdly, based on the idea of differential evolution, the teaching reform stage is proposed to avoid the algorithm falling into the local optima. Finally, the defect samples were prepared and terahertz non-destructive testing experiments were carried out. The results show that compared with the other three methods, the developed method has the best effect in eliminating local artifacts, and the two-dimensional entropy of THz images increases by 16% , 5% and 10% , respectively, and the average gradient.