The end pit is one of the important indexes for defect detection of the cylindrical lithium battery. It is very difficult to detect shallow pits automatically because the shallow pits with small contrast are easily interfered by strong noise such as bright spots and dark spots appearing randomly on metal surface. Therefore, a solution is proposed in this article. Firstly, to obtain a clear shallow pit image under a single light source angle, the six images of pit under different light source angles are collected. Secondly, the temporal averaging and outlier elimination method are used to fuse six images to obtain the datum image, and the spatial filtering method based on sliding window and Nyquist sampling theorem is utilized to weaken the interference noise with strong information intensity. Then, the average deviation is calculated according to the error analysis theory. According to the shape of pits in the gray distribution curve, the peak-tovalley difference and width ratio of concave-convex curve segment are extracted. Finally, the BP neural network is used to formulate a detection model to realize pit detection. The samples collected on site are tested, and the correct detection rate of the algorithm is 100% .