To reduce the effect of the nonuniformity of magnetic field excitation on micro-gap weld joint magneto-optical (MO) imaging, a new experimental system based on the Faraday MO effect to detect micro-gap welds (gap width less than 0.1 mm) under nonuniform magnetic field excitation was developed. Horseshoe permanent magnets were used to magnetize the weldment and establish a nonuniform magnetic field at the welding joint. MO images of the micro-gap weld joint were captured using an MO sensor under nonuniform magnetic field excitation. After analyzing the distribution characteristics of the magnetic induction intensity in the weld joint area, a characterization method for the weld zone slope was proposed. The weld zone slope could accurately determine the MO imaging effects under the nonuniform magnetic field. A model based on an error backpropagation (BP) neural network was used to predict the offset of the weld joint center at each moment, and the results performed by BP were utilized to optimize the measured value of the weld joint center. Experimental results show that it can accurately extract the position of micro-gap welds under nonuniform magnetic field excitation.