Image-, 3D- and Sensor-Based Defect and Anomaly Detection in Manufacturing Using Machine Learning and Deep Learning: A Review

Factories need to nd defects early, and learning-based inspection isspreading fast. This review asks how well it works and how well the evidenceis reported. We read 46 primary studies and 17 context papers publishedfrom late 2023 to 2026. The primary studies split into 23 individual modelsand 23 hybrid designs. Individual detectors beat their own starting model byabout two to seven points, but only on one dataset at a time. Hybrid designshelp when each part xes a different weakness. In four ablations a single partlowered the score, and in two tables a subset beat the full model. A fewbenchmarks dominate: NEU-DET appears in 10 studies and GC10-DET in 7.Only 13 of 46 studies report spread over repeated runs, and only two timetheir model on an edge board. We nd three gaps: single-dataset testing,missing deployment cost, and weak reporting of variance. The review islimited to open-access papers from a three-year window. We propose crossdatasettests, hardware-stamped timing and repeated-run reporting.