Therefore, in this paper, we will propose a data cleansing technique in the preprocessing stage to purify the dirty data, removing the noise or ambiguous feature, and rectifying the mislabel for increasing the reliability of the model and promote the performance of deep learning. A reliable dataset will bring a high-performance model, a high-performance model not only given excellent metrics also a dependable result on predict. So, data cleansing is an essential stage on the dataset`s preprocessing for giving a high-reliable and clean dataset to deep learning training.