Deep image clustering was employed to analyze the cluster structure of unlabeled image data through deep learning techniques. However, due to the absence of class labels that provide definitive information, uncertain clustering predictions may be yielded by unsupervised deep image clustering, introducing noise information that was found detrimental to performance enhancement and application development. Therefore, a clustering prediction optimization method based on alternating normalization and category-wise uniform prior was proposed to correct low confidence predictions and improve deep image clustering performance. At the same time, the method had a low degree of coupling with the model structure and training process, enabling cross-model optimization for deep image clustering frameworks. Experimental results on multiple datasets reveal that the effective clustering prediction optimization is achieved for various deep image clustering models through the approach.