Accelerating inference speed of CNN for visual inspection by filter pruning

抄録

In recent years, the industrial application of deep learning to visual inspection has been progressing along with the development of image recognition research. However, we can point out three problems with the application of deep learning to visual inspection. First, computer resources are limited when considering embedded systems. Second, tact time should be short for in situ inspection in mass-production processes. Third, high accuracy is required because overlooking a defective product can be fatal. In this study, we propose a method to build a highly accurate, lightweight, and fast inference deep learning model using a compression method called “filter-level pruning”.