Doppler weather radar is an essential tool for monitoring and warning of hazardous weather phenomena. In weather radar, achieving a longer aliasing range (ra) is crucial for surveillance, and a higher aliasing velocity (va) is also important to obtain dynamical information of storms unambiguously. However, the desire for longer ra and higher va creates a conflict because these two parameters are inversely related to the pulse repetition time (PRT). This conflict is known as the "Doppler dilemma", as ra and va cannot be improved simultaneously using a single PRT. This phenomenon is more challenging at shorter wavelengths, which means it has a more significant impact on X-band, followed by C-band and S-band. There are two main approaches to mitigating this issue. The first approach to dealias the velocity is the post-processing method. This method checks for abrupt changes from one end of the va to another, and a fold is detected when such instances are encountered. The underlying assumption is that the velocity field should be spatially continuous. This approach performs well for wide and spatially continuous storms. However, it still suffers when the storms are isolated within the radar field of view. The second approach is the waveform design method, which utilizes two or more pulse repetition times (PRTs), and the aliased velocities are found by searching for disagreement between two or more velocities observed from different PRTs. Velocity dealiasing is performed by solving a least-common-multiple problem. However, this method still has the inherent limitation of ra. The post-processing method allows the system to operate everything else, such as ground clutter filter, continuous pulse-pair processing, etc., as waveform design methods require modifications to the existing filters. Therefore, in this study, the main focus will be on the post-processing method, and the key is to detect the aliased velocity accurately, leading to the correct velocity dealiasing. The detection of aliased velocity can be compared to classification. Raw aliased velocity can be regarded as the input image, and the aliased count can be regarded as label. With advancements in technology, machine learning can be applied to image classification. Convolutional neural networks (CNNs) are widely used for image segmentation, enabling the model to output the same size as the input image. Therefore, in this study, a CNN is utilized to tackle the velocity dealiasing issue. In the training process, the input data comprises aliased velocity and the aliased count (the sign and how many times they are aliased). The best weights and the biases are determined through a fit-and-adjust process. After the training process, the performance is evaluated using unseen test data. The aliased velocity is used as input, and the output is the aliasing count. Velocity dealiasing is performed by combining the input (aliased) velocity, the aliasing count, and the known va. For evaluation, the CNN method is compared to the traditional region-based method, which is also a post-processing method in Python ARM Radar Toolkit (Py-ART). Both methods are evaluated on mostly filled precipitation and sparsely filled precipitation. Sensitivity tests are conducted on template size and the va used to optimize the CNN model to cover the X-band range coverage. This model can be used regardless of va. Both methods demonstrate similar performance on mostly filled precipitation. However, the CNN method shows better performance on sparsely filled precipitation, as it processes the entire scan at once while the region-based method only processes the limited adjacent area. The overarching goal of this study is to exploit CNN for velocity dealiasing and to achieve human-level performance. Through this process, it is expected that the labor-intensive work could be automated.