Temperature monitoring is a common requirement; the thermocouples can accurately control the temperature of rotating and moving heated object, without touching it. Uncoated metal heaters are difficult for infrared sensors to measure reliably, the reflected infrared signals can change after a heated object surface is clean, the smog is rareness after the clean heated object has been burnt, when the surface is dirty and smeary, the smog is so dense that the measurement result would be influenced. In order to measuring the metal heater accurately, the measurement noise can be reduced by the machine vision. The Self-Organizing Maps (SOM) is an efficient tool for image processing. It projects input space on prototypes of a low-dimensional regular grid. In this paper a new image process technique has been validated against U-matrix method based on Euclidean distances between input vectors and neurons weights combined with the distribution of the fixed lattices in the network. SOM, as an unsupervised neural networks, is applied to pattern recognition and image processing. By analyzing and processing of the noise signals of the image, the characteristic parameters which represent operating state of the heated object are extracted to construct characteristic vector and used to train SOM. The trained results can be used to modify the sensor testing value. A new image processing scheme based on the use of the organization property of Kohonen maps are presented in this paper, the image processing result can be correct the non-contact infrared temperature measurement.