A novel BP-GA based autofocus method for detection of circuit board components

Optical micro-inspection systems use different focusing methods depending on the inspection requirements of different scenarios. In practical industrial micro-inspection, grid samples have a wide variety of components, for example, electronic components on circuit boards and transistors on electronic chips. Changes in the surrounding environment (e.g., brightness of light, flatness of the platform, and temperature, etc.) during the inspection of these processed parts may lead to out-of-focus of the object under the microscope. Therefore, this paper proposes an autofocus algorithm to cope with the complex environment during inspection. The algorithm is based on feature vectors reflecting the external geometry of the spot and the internal energy distribution, and is combined with a back-propagation neural network with a genetic algorithm (GA) to enhance the focusing capability of the optical microscope. Preliminary numerical test results show that because of the bias problem in the focusing system, the accuracy of the neural network in calculating the defocused amount (DA) is significantly improved, despite the pitfalls of its generalization ability and the possibility of endless loops during the focusing process. In order to further solve the pitfalls of neural networks, this paper introduces a full reference image evaluation model into the optical microscope system and finally develops the autofocus software. Focusing tests using the developed software for the inspection of real components demonstrate that the introduced full reference image evaluation model not only expands the focusing distance of the inspection system, but also prevents the autofocus algorithm from falling into a dead loop. •Innovative Autofocus Algorithm: An autofocus algorithm is proposed which combines feature vector analysis and backpropagation neural networks.•High Accuracy Achievements: Defocusing amount are calculated with 98.74% accuracy and real-time performance is maintained with minimal latency.•Enhanced Generalization: The generalisation limitations of machine learning in terms of focusing accuracy are investigated and then corrected by genetic algorithms.•Full Reference Image Evaluation Model: The DISTSM model integrates texture and structural similarity to prevent autofocus from going into a dead loop.•Expanded Focusing Range: The focus range has been extended from 75 to 275 μm, while ensuring high accuracy in part inspection.