Abstract: This paper describes the use of semivariogram as a parameter for image comparison,which is a commonly used method for content-based image retrieval.The concept of semivariogram originates from spatial statistics and is one of the most important parameters for spatial data analysis.The use of this parameter for spatial pattern analysis in remotely sensed data dates back to the late 1980's,however,few literature has pointed out its potential applications to close-range photographic image analysis.This paper firstly reviews recent literature of image comparison,with the emphasis on global image comparison and distance-based image retrieval techniques.Then a practical problem,image choosing of train wheels of a moving train,is presented.It is an important task for each Chinese railway station to check the status of the wheels of a train when it is boarded in that station.The main task is to check each brake apparatus attached with each train wheel and report to relative division immediately if there are any problems.Usually,this work is done manually and mistakes are often made,especially in hot summer days when hot and humid weather makes workers light-headedness.To implement this procedure automatically is of great value.A video-recorder is used in the railway station to take photographs of wheels when the train is coming.Digital images are captured simultaneously at the rate of 6 frames per second.In order to replace manually checking the brake apparatus of each train wheel,digital image analysis methods are used to check the status of the brake apparatus.However,the first step of this procedure is to pick out useful images because there are usually about 1000 digital pictures can be captured for a single train and most of them are useless. To choose out useful images is actually a problem of image retrieval.The paper points out the difficulties of solving this problem when conventional image retrieval methods are used.The main difficulties lie in:(1) detecting features of a brake apparatus,for the image is usually moving deformed;(2) de-noiseing the moving blurring noises. Image comparison,is used to retrieve useful images.The difficulty arising here is the definition of image similarity.Many parameters are proposed to define the similarity between two images,which are proved to be invalid in this application due to the difficulties listed above.A new parameter based on semivariogram is put forward by the authors.Semivariogram is a parameter which describes not only the global structure of a data set but also the local continuity of the data set.Based on this property,a new index for image similarity is constructed and a practical program using this algorithm is developed by the authors.The results show that this algorithm has the following merits:① high correctness.All the useful images are picked out,with a few useless images mixed.The "ratio-of-picking-out" (see the definition in the paper) is around 90%;② high efficiency.It can finish the choosing procedure within 6 minutes,which completely satisfies the operational need;③ high robustness to lightening conditions.Because the calculation of semivariogram of an image is actually based on statistics,it shows high robustness when lightening conditions are changed slightly.This paper not only extends the use of semivariogram from remotely sensed images to close-range photographic images,but also puts forward a new parameter to describe image similarity.