SPLConv: Channel-Separated Convolution Using Large Convolution Kernels
DOI:
https://doi.org/10.56028/aetr.9.1.188.2024Keywords:
CNN; Defect segmentation; Model compression.Abstract
In modern society, chips play a crucial role in various aspects of everyday life. Nevertheless, the diverse chip manufacturing processes can introduce defects, making the detection of faulty chips imperative. Convolutional neural networks (CNNs) are extensively employed in computer vision due to their ability to deliver high-performance results. However, this comes at the expense of significant computational resources. Current research focuses on employing pruning or quantization techniques to minimize the size of model parameters and investigate the development of lightweight models. This paper proposes a large depth-wise convolutional module called split-based large convolutional (SPLConv), which is devised and optimized from the perspective of split-based convolutional operations (SPConv). SPLConv not only disregards redundant features in the convolutional layer but also enhances spatial information capture, thereby mitigating excessive computational overhead. Experiment demonstrate that SPLConv effectively reduces complexity and computational cost while delivering commendable performance.