Aiming at failure detection problems on subtle region caused by saliency differences of detected target in local region, under the framework of Bayesian theory, the author proposed a novel salient region detection method based on cellular automata multi-scale optimization. Firstly, the prior information about dark channel was integrated with regional contrast to separately construct original salient maps in five superpixel scale spaces on the same picture; and then the cellular automata was used to establish a dynamic updating mechanism and impact factor matrix and confidence matrix were applied to optimize influences of each cellular in next state. As a result, the saliency values of all cells will be renovated simultaneously according to the proposed updating rule, and five optimized salient maps were obtained; finally, under the framework of fusion algorithm in Bayesian theory, the final saliency map was obtained. The experiment on two standard image datasets with different complexity was conducted, and experimental result indicates that the performance of proposed algorithm is superior to other ten existing salient region detection algorithms both in visual effect and in objective quantitative comparison. Especially on the most challenging DUT-OMRON data base, the aggregative indicator F-measure value of proposed algorithm is 0.631 4, and mean absolute error (MAE) is 0.132 5 and ROC area under the curve (AUC) is 0.892 8, indicating that the algorithm has higher accuracy and robustness.