By extracting the iron cap and disc surface temperature information of the disc suspended porcelain insulator string in the singleframe infrared image, the relative temperature difference is considered as a criterion to diagnose the deterioration state, which is an efficient and accurate method for online automatic monitoring of the insulator string state. In order to accurately extract the temperature information, this paper proposes an algorithm that combines the characteristics of the insulator image and deep learning to accurately identify the iron cap and disc surface area of the porcelain insulator string in the infrared image. The algorithm uses a large number of different parts of insulator images as sample data sets, and is trained by selfconstructed CNN to form three classifiers. Then it uses the classifiers to identify in the corrected insulator string region image and finally utilizes different colors to mark in the original infrared image. It is shown that this algorithm can obtain excellent recognition results of cap and disc area for insulator strings of different voltage levels and different disc types.