Blood glucose levels are an important indicator of human health. Blood glucose monitoring is crucial in the treatment of diabetes and other diseases. However, traditional invasive blood glucose measurement methods carry risks such as wound infection and pain. Therefore, this paper proposes a non-invasive blood glucose detection algorithm based on facial infrared thermography combined with an improved MobileNet-V3. First, histogram equalization is used to enhance the contrast of the infrared images. Then, the activation functions and attention mechanism modules of MobileNet-V3 are improved to enhance the model’s expressive power, comprehensively considering features from both the channel and spatial dimensions. Experimental results show that the results of the integrated improved model fall into zone A in the Clarke error grid analysis is 91.573%, which is an improvement of 9.253% over the pre-improvement period.