The introduction of sparse code multiple access (SCMA) is driven by the high expectations for future cellular systems. In traditional SCMA receivers, the message passing algorithm (MPA) is commonly employed for received-signal decoding. However, the high computational complexity of the MPA falls short in meeting the low latency requirements of modern communications. Deep learning (DL) has been proven to be applicable in the field of signal detection with low computational complexity and low bit error rate (BER). To enhance the decoding performance of SCMA systems, we present a novel approach that replaces the complex operation of separating codewords of individual sub-users from overlapping codewords using classifying images and is suitable for efficient handling by lightweight graph neural networks. The eigenvalues of training images contain crucial information, such as the amplitude and phase of received signals, as well as channel characteristics. Simulation results show that our proposed scheme has better BER performance and lower computational complexity than other previous SCMA decoding strategies.