Compression-Based Data Augmentation for CNN Generalization

Nowadays, deep learning is widely exploited in various fields due to its ability to solve complex problems. These networks have proven their efficiency compared to classical machine learning methods in several recent applications. Machine vision, as an innovative technology, represents a major element of the industrial transformation and is currently replacing human vision, which is outpaced by the speed and complexity of the tasks in most manufacturing processes. However, optimizing latency is an important challenge for integrating machine vision into real-world use cases. In this context, compression of collected images before they are transmitted and processed is crucial to save bandwidth and energy, and enhance latency in vision applications. Nevertheless, the degradation of image quality resulting from compression affects the performance of convolutional neural networks (CNNs) and reduces the accuracy of the results. In this paper, a compression-based data augmentation method is proposed to improve the classification performance of CNNs and generalize the models when tested on poor compression qualities. Three different models were trained and tested with images from the surface defect database. The obtained results in the performed experiments reveal that the compression-based data augmentation significantly increases the classification precision of CNNs, and improves the generalization of the models when tested on different compression qualities.