To find unknown point gamma sources in large areas, we proposed a source direction estimation method based on self-shielding profiles of 3D multi-pixel array detectors. A convolutional neural network model was investigated to estimate the direction of the gamma-ray source in 4π directions to the detector. The estimation of source direction for a single 137Cs point source was demonstrated both computationally and experimentally using a multi-pixel CdTe array detector. The validation results showed that 59.2% of the estimated results were within 10 degrees of the true incident direction.
