Character Identification for Integrated Circuit Components on Printed Circuit Boards Using Deep Learning

Conventional defect detection and recognition algorithms for integrated circuit (IC) components, used in printed circuit boards (PCBs), are primarily based on automated optical inspection (AOI). These techniques typically extract image features using image processing models, which are heavily dependent on visual cues and can thus be inaccurate. To address these issues, we propose the use of a deep convolutional neural network, an improved LeNet-5 structure called ICChaNet with a deeper structure and more complex parameters. First, the algorithm for character extraction is implemented and used to generate character samples from IC component images. This process includes grayscaling, binarization, and contour extraction, applied to a set of IC component images to establish 53 categories of characters (letters and number with similar symbols removed). Three critical hyperparameters, including learning rate, pooling strategy, and optimization strategy, are then optimized through a comparative analysis. The effect of different network architectures on model efficiency is also investigated by varying the layer depth. A combination classifier strategy is subsequently added to the fully connected layers to strengthen the feature expression in the corresponding classes. Finally, the performance of the proposed ICChaNet model is verified through a comparison with popular CNNs. Experimental results demonstrate that ICChaNet has achieved a classification accuracy of 98.5% and a test average accuracy precision of 97.73%. These results suggest the proposed ICChaNet model is a promising approach for automated character identification.