Article Preview
TopIntroduction
According to industry forecasts, the PCB manufacturing industry is forecast to grow $89.7 billion by 2024. The work involved with producing PCBs is extremely precise and complex and there is an enormous demand for PCBs, so PCB manufacturers must produce a large quantity at a time (Hassanin et al. 2019; Seddik 2014; Zaibi 2021). These circumstances raise a question about how these manufacturers can ensure the quality of their PCBs while producing large quantities. Automated inspection systems help to solve the aforementioned problem, where a lot of research done in recent years to detect defects in PCBs. However, the results of these researches were ineffective and did not allow for the detection of tiny defects. Recently, the PCB manufacturers use different image processing methods such as image subtraction, template matching, etc., but are still unable to keep up with quality inspection processes (Wang et al. 2018; Alelaumi et al. 2020). So, the combination of traditional image processing models and machine learning models is required to fill this gap. Recent advances in deep learning have allowed developers to obtain more generalized solutions in computer vision. Convolution neural networks (CNNs) in particular have proven highly effective at solving recognition and detection problems. In contrast to conjugating techniques for obtaining features, CNNs can learn features from images automatically and can operate without learning different techniques (Dai et al. 2019; Zhang et al. 2022; Gaidhane et al. 2018; Kim et al. 2021). This research chose to work with the Leaky Rectified Linear unit (ReLu) activation function on the LeNet-5 (Leaky-LeNet-5) network. After testing the proposed network on 6 categories of PCB defect detection (PCB-DD) cases in still images, SURF and LeNet-5 are supporting to achieve effective results. The research across the globe modified the LeNet-5 architecture in various ways, which is an important aspect of this research. Improvements to classic LeNet-5 resulted in excellent and even perfect accuracy, lightness, and parameter reduction compared to the recent year model (Adibhatla et al. 2020; Shen et al. 2020; Miao et al. 2021).
The main contributions of this study are listed as follows:
- •
The main goal of this research is to develop a CNN-based recognition system for PCB. The Leaky ReLu activation function is employed to improve the backpropagation during zero gradient value.
- •
SURF is employed to specifically find the location of the defective portion instantly. These additive feature maps will improve the performance of the LeNet-5.
- •
The proposed architecture was trained to find the multiple defects in a single test file and 2953 images are used for validation purposes.
This paper is organized as follows: Few existing research papers related to “PCB defect detection and classification” are reviewed in Section 2. The detailed explanation and the experimental results of the proposed SURF-Leaky LeNet-5 model are denoted in Sections 3 and 4. The conclusion of the present work is stated in Section 5.
TopIn recent years, several techniques have been used to detect the defect during the PCB manufacturing process. The existing literature is as follows:
Dimitriou et al. 2020 applied a deep learning method for the detection of defects in PCB and its alternative method based on deep neural networks could simulate changes in a monitored object's 3D structure based on the 3D measurement history in the study. In particular, 3DCNNs were proposed as a method for predicting future events about sub-optimal performance associated with geometric variations in manufacturing parameters. The developed model has limitations of successive 3D scans of automated registration and is difficult for complicated 3D objects.