For nano defect inspection, it is required to detect weak scattered light. Ghost imaging (GI) is a known imaging method for its high sensitivity and high robustness. To reduce the measurement time, the defects inspection method combined GI with deep learning has been proposed. However, robustness is sacrificed to reduce the measurement time. To overcome this problem, we propose the GI with complementary correlation computation using deep learning. This report shows that, in noisy environments, it improves accuracy using weighted illumination patterns as input for CNN by numerical analysis.