An optical communication method using deep learning in a code-division multiple access (CDMA) scheme may include obtaining an image of an optical signal from a light source and applying a first model trained to identify a location of the light source to detect the light source in the image. The method may also include detecting a preamble of an on-off keying (OOK) modulated signal from a pulse signal generated based on the identified light source using a second model trained to detect preambles in received signals, to decode the OOK modulated signal for a first user device by applying a first pseudo-random noise (PN) code. The method may further include determining a start position of an orthogonal frequency division multiplexing (OFDM) frame in an OFDM modulated signal from the pulse signal, to decode the OFDM modulated signal by applying a first Walsh code.