Colorimetry often suffers from deficiency in quantitative determination, susceptibility to ambient illuminance, and low sensitivity and visual resolution to tiny color changes. To offset these deficiencies, we incorporate deep machine learning into colorimetry by introducing a convolutional neural network (CNN) with powerful parallel processing, self-organization, and self-learning capabilities. As a proof of concept, a plasmonic nanosensor is proposed for the colorimetric detection of glucose by coupling Benedict’s reagent with gold nanoparticles (AuNPs), which relies on the assemble of AuNPs into dendritic nanochains by Cu₂O. The distinct difference of refractive index between Cu₂O and Au and the localized surface plasmon resonance coupling effect among AuNPs leads to a broad spectral shift as well as abundant color changes, thereby providing sufficient data for self-learning enabled by machine learning. The CNN is then used to fully diversify the learning and training of the images from standard samples under different ambient conditions and to obtain a classifier that can not only recognize tiny color changes that are imperceptible to human eyes, but also exhibit high accuracy and excellent anti-environmental interference capability. This classifier is then compiled as an application (APP) and implanted into a smartphone with Android environment. 306 clinical urine samples were detected using the proposed method and the results showed a satisfactory correlation (87.6%) with that of a standard blood glucose test method. More importantly, this method can be generalized to other applications in colorimetry, and more broadly, in other scientific domains that involve image analysis and quantification.