Aiming at the low efficiency of manual detection in the detection of water channel defects of engine cylinder block and the poor generalization ability of traditional machine vision in manual design features, an improved water channel defect detection model of cylinder block based on fast RCNN network is proposed. Restnet50 is selected as the feature extraction network of waterway defects, and feature pyramid network (FPN) is introduced to improve the detection ability of small defects; The anchor box is optimized by k-means++ clustering algorithm to improve the positioning of the target box. Experiments show that the map of the improved network in the engine cylinder block waterway defect data set reaches 88.74%, the accuracy is increased by 4% compared with the original fast RCNN, and the recall rate is increased by 1.67%. The recognition and detection effect in real samples is good, and it can be effectively used for the engine cylinder block waterway defect detection.