This paper presents a finger motion estimation based on sparse multi-channel surface electromyography (sEMG) signals using a convolutional neural network (CNN). Although classification with CNNs has achieved high accuracy in gesture recognition, the most cases use a high-density sEMG as the signal acquisition method, which is problematic because this requires many sensors for measuring sEMG signals, resulting in high costs. We therefore propose estimating the finger motion with a sparse multi-channel sEMG method using ring-shaped sensors. The finger motion estimation is performed by classifying images generated from the amplitude variations of sEMG signals, and the image classification is achieved with a simple CNN model featuring two pairs of convolutional and pooling layers and two fully connected layers. Experimental results showed that the test accuracy reached 90% in classifying sEMG signals into four types: thumb opened, thumb closed, fingers (excluding thumb) opened, and fingers (excluding thumb) closed.