Algorithm for Size Statistic and Clustering of Block Image Sequences

The statistic algorithm for image processing, block clustering and block size is proposed according to the block image sequences. For complicated and regular distribution block image. Firstly, we get the binary picture, and calculate the ratio of the edge and ground and the maximum distance between the two edges by the pixel scanning method, then we can get the data sample of the size and block size distribution at the same time. Secondly, using the improved nearest neighbor algorithm[5], we get the block size classification, and the percentage of each regional size in the whole area. Computer emulate result proved that this method meets the demands of the real-time image processing better for its advantage as follows: the small calculation, the high precision. The algorithm will lay a solid foundation for the operating mode analysis and automatic control.


Introduction
There need a real-time observation of the melting or calcination dynamic process of some material, in order to take analysis or corresponding parameter control.For instance, we need immediate understanding the quantity and size of the clinker caking in the cement rotary kiln calcination process, Because the quality of cement products not only related to the changes of material sintering temperature, but also related to the quantity and size of the clinker caking, and the quantity and size of the clinker caking mainly depends on the raw material ratio and appropriate mineral composition and petrographic structure .However, these processes especially like the cement rotary kiln calcination process is a very complicated production process, it is hard to establish mathematical model, and its particle size detection problems are also difficult to solve.We can use popular soft measurement technology [1] and computer vision detection technology to take block sintering images [2], which including abundant production condition information-the ratio of the edge and ground, block distribution state, then by using the image processing technology, based on identification principle to get characteristic quantity [3][4], which play a decisive role in production condition.On the base of the accumulation of a great number of statistic data, we can establish quality prediction model, and then realize the quality optimization control of the process.
The Purpose and motivation of the statistic algorithm for the block image sequences clustering and block size distribution is obvious.On the base of the accumulation of the distribution law of the block, the author acquire the simulation image of in the furnace caking with CCD in laboratory (as shown in Fig. 1), and get the algorithm to solve block size distribution after repeated research and argumentat sample of ground and using the im demands o the high sp

We can assumed th
We need  Also can again with the Y axis scanned, for accurate, if time allows, can two axis scanned, will get size of combination.We can also scanning by Y axis, if time allows, also can scanning by two axis in order to be accurate, and combined the obtained size.

Block Clustering
Literature [5] demonstrated many advantages of the new clustering algorithm in detail compared to other modes clustering algorithm in detail.This paper improved on the basis of this algorithm.
Firstly, we determine a few reasonable rules to be followed before a large number of sample data is clustered into a limited clusters: Set clustering center at the area where the data is more dense; data within a neighborhood of the clustering center is categorized as one cluster, called the hinterland of this clustering center; The closeness degree between clustering center and the data within its hinterland reflect the quality of clustering results, the closer between clustering center and the data within its hinterland, the better the quality of clustering results are; The treatment process of calculate the clustering center and the classification repeatedly can make the clustering results optimized gradually.
We use the generalized distance to indicate the degree of closeness between the data, which can be defined with any norm, here we choose Euclidean norm, and the distance is:


, each sample data contains n components.Now the clustering center is: Where x ) (m j is all sample in the class which clustered with m c as the center, J is the sample number of the m th clustering.m used to count each clustering, Assuming there are a total of k clusters, k m   1 .Separately calculated the distance between sample x and all clustering centers , … . Then x belongs to clustering p c .We rule all sample belong to clustering m c constitute the hinterland of clustering m c , denoted by.m IC .Assume the total number of sample as E , the sample number the m th cluster as m E , so that the rate is defined as follows: E E 4 The total rate is defined as follows: 5 For the cluster centered on m c , Define its devious ) (m D as follows: reflects the devious degree of the m th clustering data, or in other words compact degree.The total devious D is defined as follows: 7 The smaller the total devious D is, the more compact the cluster is, and thus the better the quality of clustering result is.The clustering process: Firstly, select the points distance minimum between each other as the clustering center, and classified the samples within its certain range into its hinterland.Then choose the points distance minimum between each other as the clustering center from residual samples, and determine its hinterland.Samples have already been processed will not be considered when continue to choose clustering center and the classification.The points have a larger distance between each other will not get together.Repeat the above process until all samples are clustered.This clustering method can avoid the cluster focus on local area, The samples have a larger distance between each other will be forced to get together when the number of cluster is limited, which make the total devious very large, so that largely avoided the cluster fall into local solution.We need to make adjustments to further reduce the total devious after the initial clustering is finished, and maximize the total rate.Firstly, calculate the center of each cluster, secondly, calculate the distance between all samples and the center of each cluster and reclassified all samples according to the principle of formula (3).Adjust the clustering repeatedly until the total devious no longer decreased significantly and the total rate is maximum.

The experimental results
We get a image sequence by shooting continuously in a certain time interval.Process each picture in this sequence respectively according to the above.We conducted field simulation in the laboratory.Material stones replaced by standard concrete experiment stones, each distribution rate of the stones whose size is around the 50mm, 35mm and 20mm is 20%, 30% and 50%.We change the position of different size block properly.theadopted 20 picture constitute a set of image sequence.The extraction results of the size in 20 pictures are basically similar.We classify the stones into three clusters (can also divide them into several kinds, or take treat accordingly treatment after segmenting the images are segmented when there are longitudinal and transverse block partition at the same time).The statistical result is that stones size is 42.3mm, 37mm and 16mm respectively.The rate of the major medium and small stones in the whole area is that: 50mm (40-55mm) caking quantity is 24.3%, 35mm (25-40mm) caking quantity is 26.2%, 20mm (15-25mm) caking quantity accounts for 45.1%.The results are satisfactory compared with the actual results.Other depth algorithm needs further improvement.
Fig.2.Size and D om the orig attice distan Then we can n the j th ed