Arrival-time picking is an essential step in seismic processing and imaging. The explosion of seismic data volume requires automated arrival-time picking in a faster and more reliable way than existing methods. We have treated arrival-time picking as a binary image segmentation problem and used an improved pixel-wise convolutional network to pick arrival times automatically. Incorporating continuous spatial information in training enables us to preserve the arrival-time correlation between nearby traces, thus helping to reduce the risk of picking outliers that are common in a traditional trace-by-trace picking method. To train the network, we first convert seismic traces into gray-scale images. Image pixels before manually picked arrival times are labeled with zeros, and those after are tagged with ones. After training and validation, the network automatically learns representative features and generates a probability map to predict the arrival time. We apply the network to a field microseismic data set that was not used for training or validation to test the performance of the method. Then, we analyze the effects of training data volume and signal-to-noise ratio on our autopicking method. We also find the difference between 1D and 2D training data with borehole seismic data. Microseismic and borehole seismic data indicate the proposed network can improve efficiency and accuracy over traditional automated picking methods.