Abstract: To solve the existing problems of low efficiency, inadequate accuracy, and a limited quantity of solder joint image samples in current circuit solder joint defect detection methods, a methodology was proposed based on metric learning for the expeditious identification of solder joint defects. Firstly, industrial cameras with telecentric lenses were arranged to capture solder joint images. And then, extracting the inherent feature of the solder joint images, a cross-point detection method was devised to segment the images of the welding units, constructing a dataset with instances of solder joint defects. Building upon this foundation, a scheme was developed to integrate global and local feature extraction methodologies of solder joint images, amalgamating the two distinctive features of solder joints. Moreover, the improvements were carried out for the attention mechanism, incorporating it into the global feature extraction module. Finally, the detection of solder joint defects was realized. The detection results show that the accuracy rate can reach up to 98.4%, meeting the actual production requirements of solder joint defect detection.