<p>This study develops an automated optical inspection system for detecting surface defects in Micro-LED chips. The system integrates an industrial camera, a microscope, an objective lens, and a linear slide to automate image collection. It employs a Bayesian optimization (BO) method and the Support Vector Data Description (SVDD) algorithm to identify four defect types: holes, fractures, missing chips, and dislocations. A decision tree method is used to achieve chip classification, which in turn facilitates wafer classification. Using industry-standard Micro-LED data, we demonstrated that the classifier can effectively detect most defect types. Experimental results confirm that the proposed inspection system can accurately detect Micro-LED pixels, defective pixels, defect types, and defect coordinates. This enables production line personnel to implement process rollbacks and reduce defects caused by mass transfer. It also provides repair coordinates for subsequent mass repair processes. The study addresses the industry’s challenge of imbalanced data, where standard samples significantly outnumber abnormal ones. It leverages SVDD and BO to enhance classifier training. Template matching is used to segment images and define regions of interest, improving calculation speed by over five times. A comparative analysis between SVDD and Support Vector Machine (SVM) reveals that while both models achieve similar classification rates on standard samples, SVDD outperforms SVM on defective samples, achieving a 100% negative classification rate compared to SVM’s 0%. This approach effectively mitigates poor classification performance caused by data imbalance, bias, or overfitting, providing a robust solution for detecting Micro-LED chip defects.</p>