Targets Recognition Based on Feature Extraction with Small Convolutional Networks for Ultrasonic NDT

In previous image defect recognition algorithms, machine learning methods were used for feature optimization to identify image defects. Although good recognition accuracy could not be achieved for ultrasonic NDT data sets due to the speckle noises, further improvement to reach the desired recognition results was challenging. Traditional algorithm classifiers’ recognition performance depends on the function of the extracted features. Hand-crafted features require analysis specific to defects, extensive experiential design, and a deep understanding of image defects. To avoid the influence of multiple processing procedures and human factors, this paper primarily combines popular convolutional neural network recognition and classification algorithms to construct a small convolutional network model. This forms an integrated feature extraction and recognition approach, achieving an end-to-end recognition model. Furtherly conducts an analysis of network structure adjustments, convolutional kernels, and over-fitting solution for ultrasonic NDT images.