Expert training: Enhancing AI resilience to image coding artifacts

Abstract

In the Machine-to-Machine (M2M) transmission context, there is a great need to reduce the amount of transmitted information using lossy compression. However, commonly used image compression methods are designed for human perception, not for Artificial Intelligence (AI) algorithms performances. It is known that these compression distortions affect many deep learning based architectures on several computer vision tasks. In this paper, we focus on the classification task and propose a new approach, named expert training, to enhance Convolutional Neural Networks (CNNs) resilience to compression distortions. We validated our approach using MnasNet and ResNet50 architectures, against image compression distortions introduced by three commonly used methods (JPEG, J2K and BPG), on the ImageNet dataset. The results showed a better robustness of these two architectures against the tested coding artifacts using the proposed expert training approach. Our code is publicly available at https://github.com/albmarie/expert training.

Video Coding for Machines: Large-Scale Evaluation of Deep Neural Networks Robustness to Compression Artifacts for Semantic Segmentation

“…Convex hull of VTM-12.0 with ALL-Intra configuration is used as an anchor, where QP ∈ 22, 27, 32, 37, 42, 47 and downsampling factor δ ∈ {0.25, 0.5, 0.75, 1.0} are considered. [15], [28]. As shown by Figure 2, removing (IV) chrominance information is beneficial for JPEG at lower bitrates.…”

“…This is because some studies only compare themselves to HEVC test Model (HM) or VVC Test Model (VTM) with few Quantization Parameter (QP) after proposing a new method to reach better tradeoffs between vision task performance and bitrate [14]- [17], [19], [25], [26]. Some papers also evaluate DNN resilience to JPEG/JPEG2000 compression [10], [12], [15], [18], [21], [24], [27], [28], [30], AVC [1], [31] or auto-encoders [18], [26], [27], but no paper consider all mentioned image and video codec generations in a unified framework (II). Note that older codecs such as JPEG or AVC achieve lower trade-offs between bitrate and vision task performance, but their low-complexity compared to modern codecs makes them more suitable to some applications using low-power devices [31], especially when hardware implementation of AVC encoders is still widespread nowadays.…”

“…Few papers attempted to (I) use compressed data at training time in order to improve DNN resilience to compression artifacts [15], [20], [22], [24], [27], [28], [34]. All studies converge to show that adding compressed data at training time allows to reach much higher trade-offs between bitrate and vision task performance.…”

“…All studies converge to show that adding compressed data at training time allows to reach much higher trade-offs between bitrate and vision task performance. While training and evaluation with the same encoding configuration allows to reach the best trade-offs, training and evaluating with different codec is still beneficial for DNN resilience [15], [27], [28].…”