Design Of Specimen Crack Image Recognition System On Temperature-Stress Test Machine

At present, infrastructure construction as well as the real estate industry is occupying the peak in China, so the concrete usage is very large. On the other hand, the problem of concrete crack is quite universal. Therefore this problem appears extremely serious and urgent. The evaluation of concrete crack resistance is the basis which the engineering design, the construction and selecting the raw material are according to. It can effectively guarantee the concrete satisfies the stipulation of the service life in the different environment. Now the problem also is one of hot topics in lots of countries' concrete research area.In this dissertation, specimen crack image recognition system on temperature-stress test machine is researched. Having implemented a method combining theory with simulation, it mainly describes crack image preprocessing, crack image segmentation, a BP neural network based crack image classifier and so on.It is concisely introduced the scheme of specimen crack image recognition system, main hardware structure and software architecture. In view of the concrete crack image characteristic, it emphatically discussed the recognition system three core issues.Firstly, it particularizes image preprocessing technology for reducing noise of concrete crack image. On the base of anglicizing the noise of specimen crack image and the characteristic of the noise , it propose the scheme that a hybrid filter will be used to reduce noise and introduce the principle of a hybrid filter.Secondly, after studied crack image segmentation and the characteristic of unimodal histogram of concrete crack image, it concisely introduces the histogram revision technology based iterative and pruning algorithm to carry on image segmentation.Thirdly, in view of surface crack image classification, the dissertation describes a method for crack image classification based on BP neural network, which profits from the domestic and foreign research. First of All, The system obtain characteristic crack image by divided the binary image to sub images. Finally three representative features are picked up, difference sum of the characteristic crack image projection in horizontal (Ph) and vertical direction (Pv), as well as the pixel number of crack sub image (Num). Selecting these three features as inputs, it designs a three-layer BP neural network. Outputs of the network are 5 types of crack: transverse, longitudinal, combination of transverse and longitudinal, alligator and non-distress. An artificial set of crack image training data is used to train the BP neural network, and the validation of the BP neural network is tested. The result has shown that the performance of this neural network can perfectly match the industry usage. The accuracy of the BP neural network well achieves a high level.