Improvement of Deep Learning Models Using Retinal Filter: A Systematic Evaluation of the Effect of Gaussian Filtering With a Focus on Industrial Inspection Data

The range of products manufactured in the industry is vast, as are the anomalies that may occur during the manufacturing process. Some companies have initiated the automation of the quality control process and the implementation of modern deep learning methodologies to promptly identify defective objects. In addition to architectural optimizations within the deep learning model, the utilization of pre-processing filters can also be employed to help improve recognition performance. The Gaussian filter serves as an illustrative example, exhibiting functionality that parallels aspects of the human visual processing system. Although some authors have successfully applied this filter, a comprehensive study assessing its actual impact on a large number of objects and the extent to which various performance indicators are affected remains pending, which is the primary objective of this study. To this end, a variety of architectural approaches are employed, including Xception, InceptionV3, ResNet50V2, VGG19, and VGG16. In addition to hyperparameter tuning, all architectures utilize transfer learning. The results demonstrate that the implementation of an innovative Gaussian filter approach enhances the balanced accuracy in at least one architecture for 13 out of 15 product categories. Furthermore, the filtering approach positively influences various performance indicators across the majority of categories. It can be concluded that the Gaussian filter is often an effective technique for enhancing model performance across various product categories, making it a valuable and efficient tool for industrial defect detection in quality control applications. This study provides an overview of the results achieved and other key performance indicators for all the models used.