Long-term and frequent blood glucose detection by requiring finger-pick blood become unrealistic. An appropriate non-invasive detection system is thus highly desirable to deal better with it. In this paper, a non-invasive and intelligent dual-sensing system is reported. The feasibility of proposed system has been verified by glucose solution, animal serum and human trials. In the in-vivo experiments, detection signal has a high correlation (r = 0.96) with blood glucose level. An improved convolution neural network (cascade CNN) is purposed to be employed for estimation of blood glucose level (BGL). For the estimation results of BGL, the root mean squared error (RMSE) of 7.3217 mg/dL and mean absolute relative difference (MARD) of 4.7209% were achieved. The estimated results were also 100% fallen into the clinically acceptable zones of the Clarke error grid analysis, which indicated that proposed system could be potentially used for clinical measuring.