Considering that different types of defects have different failure modes and effects on pipeline safety, recognizing the type of detected flaw is crucial for further evaluation [
28]. To improve defect recognition accuracy and realize in-service inspection, automatic defect recognition technique was developed [
29]. Xie et al. [
30] applied the principle of spatial compound imaging to reduce the noise in ultrasonic image, and consequently improved the image quality and defect detection ability. Rostami and Razavi [
31] proposed a combined algorithm through various image processing and mathematical morphology to improve the raw ultrasonic images for observation and accurate analysis. Nevertheless, both the methods failed to differentiate defects from each other, and defect recognition still depends on experienced technicians. Huang et al. [
32] summed the pixel grey-values in B-scan images of EF joint horizontally, normalized the value to one-dimensional signal, and analysed the signal by wavelet transform to facilitate automatic cold welding defect detection. Long et al. [
33] conducted principal component analysis and regression analysis on eight features extracted from signal wave of voids and established the relation between the features and defect information. TWI [
34] developed an automatic defect recognition (ADR) software for EF weld inspection, including three main steps in the algorithm: Detection of heating wire zone, determination of wire lines, and determination of defect maps. Based on the ADR software, defect indication can be completed, but defect classification and quantification required further analysis by technicians. Hou et al. [
21] proposed a defect recognition method for EF joints based on pattern recognition principle, mainly including feature extraction and selection, defect classification, and defect quantification, as shown in Fig. 4. A series of software [
35,
36] were further developed to perform automatic weld defect identification for phased array ultrasonic inspection, and the accuracy rate of single defect recognition reached 90%.