Depression has become an important public health problem in recent years because the probability of a depressive episode in a person's entire life is generally between 18-20%. Neuroimaging techniques investigate diagnostic biomarkers in depression disorders and support traditional communication-based diagnosis in psychiatry. The quality of the brain images used in functional MRI (fMRI), and the design of decision support systems using these images are essential for accurate diagnosis. The Gaussian smoothing for fMRI preprocessing blurs the image for statistical analysis but is inadequate because image detail is lost during filtering, leading to poor classification results. We argue that the weighted-3 Dimensional-Discrete Wavelet Transform (weighted-3D-DWT) denoising approach instead of Gaussian smoothing for task-based fMRI. The activation maps show differences in intensity values in the cluster size of voxels in the mood-related regions between patients and control subjects (p<0.05). Thus, we classify depression disorders using a machine learning approach and improve the classification accuracy using weighted-3D-DWT. The classification results show that weighted-3D- DWT with Random Forest and 10-fold cross-validation achieves 96.4% accuracy, while Gaussian Smoothing with a Support Vector Machine achieves 83.9% classification accuracy. Classification accuracy increases to 97.3% when 30 components are used with principal component analysis. Our results show that an fMRI experiment with visual stimuli that can aid the diagnosis of depression provides significant classification accuracy using weighted-3D-DWT.