Cracks of girth welds of longdistance oil and gas pipeline bring extremely harm to the pipeline safety, and most of accidents caused by pipeline defects occurred at pipeline welds. So far non destructive testing (NDT) is a common method for predicting potential risk and ensuring the safe operation of pipeline, but traditional NDT methods can’t effectively identify the defects lying in girth welds or other complex surface. In order to overcome the disadvantages of traditional methods, an embedded eddy current testing system is presented based on image processing and neural network. Hough transform and contour extraction are used to extract the features from 2D impedance image composed of eddy current signals. Features with good classification characteristics are selected by the within class scatter matrix to train neural network based on FPGA speeding up. Automatic cracks of girth welds classification and identification is realized with a heavy weld noise floor. Experimental results show that this system can effectively identify the defect signals lying in weld of the cylinder or other complex surface. The accuracy of the system is as high as 92%, and has lower power consumption and smaller size, which is suitable for pipeline inner inspection.