Herpes zoster (HZ) is a common cutaneous disease affecting one out of five people; hence, early diagnosis of HZ is crucial as it can progress to chronic pain syndrome if antiviral treatment is not provided within 72 hr. Mobile diagnosis of HZ with the assistance of artificial intelligence can prevent neuropathic pain while reducing clinicians’ fatigue and diagnosis cost. However, the clinical images captured from daily mobile devices likely contain visual corruptions, such as motion blur and noise, which can easily mislead the automated system. Hence, this paper aims to train a robust and mobile deep neural network (DNN) that can distinguish HZ from other skin diseases using user-submitted images. To enhance robustness while retaining low computational cost, we propose a knowledge distillation from ensemble via curriculum training (KDE-CT) wherein a student network learns from a stronger teacher network progressively. We established skin diseases dataset for HZ diagnosis and evaluated the robustness against 75 types of corruption. A total of 13 different DNNs was evaluated on both clean and corrupted images. The experiment result shows that the proposed KDE-CT significantly improves corruption robustness when compared with other methods. Our trained MobileNetV3-Small achieved more robust performance (93.5% overall accuracy, 67.6 mean corruption error) than the DNN ensemble with smaller computation (549x fewer multiply-and-accumulate operations), which makes it suitable for mobile skin lesion analysis.

A simple, novel, rapid, and non-destructive spectroscopic method that employs the deep spectral network for beef-freshness classification was developed. The deep-learning-based model classified beef freshness by learning myoglobin information and reflectance spectra over different freshness states. The reflectance spectra (480-920 nm) were measured from 78 beef samples for 17 days, and the datasets were sorted into three freshness classes based on their pH values. Myoglobin information showed statistically significant differences depending on the freshness; consequently, it was utilized as a crucial parameter for classification. The model exhibited improved performance when the reflectance spectra were combined with the myoglobin information. The accuracy of the proposed model improved to 91.9%, whereas that of the single-spectra model was 83.6%. Further, a high value for the area under the receiver operating characteristic curve (0.958) was recorded. This study provides a basis for future studies on the investigation of myoglobin information associated with meat freshness.

We propose the implementation of transfer learning from natural images to audio-based images using self-supervised learning schemes. Through self-supervised learning, convolutional neural networks (CNNs) can learn the general representation of natural images without labels. In this study, a convolutional neural network was pre-trained with natural images (ImageNet) via self-supervised learning; subsequently, it was fine-tuned on the target audio samples. Pre-training with the self-supervised learning scheme significantly improved the sound classification performance when validated on the following benchmarks: ESC-50, UrbanSound8k, and GTZAN. The network pre-trained via self-supervised learning achieved a similar level of accuracy as those pre-trained using a supervised method that require labels. Therefore, we demonstrated that transfer learning from natural images contributes to improvements in audio-related tasks, and self-supervised learning with natural images is adequate for pre-training scheme in terms of simplicity and effectiveness.

A haptic sensor coupled to a gamepad or headset is frequently used to enhance the sense of immersion for game players. However, providing haptic feedback for appropriate sound effects involves specialized audio engineering techniques to identify target sounds that vary according to the game. We propose a deep learning-based method for sound event detection (SED) to determine the optimal timing of haptic feedback in extremely noisy environments. To accomplish this, we introduce the BattleSound dataset, which contains a large volume of game sound recordings of game effects and other distracting sounds, including voice chats from a PlayerUnknown’s Battlegrounds (PUBG) game. Given the highly noisy and distracting nature of war-game environments, we set the annotation interval to 0.5 s, which is significantly shorter than the existing benchmarks for SED, to increase the likelihood that the annotated label contains sound from a single source. As a baseline, we adopt mobile-sized deep learning models to perform two tasks: weapon sound event detection (WSED) and voice chat activity detection (VCAD). The accuracy of the models trained on BattleSound was greater than 90% for both tasks; thus, BattleSound enables real-time game sound recognition in noisy environments via deep learning. In addition, we demonstrated that performance degraded significantly when the annotation interval was greater than 0.5 s, indicating that the BattleSound with short annotation intervals is advantageous for SED applications that demand real-time inferences.

In this study, we introduce a feature knowledge distillation framework to improve low-resolution (LR) face recognition performance using knowledge obtained from high-resolution (HR) images. The proposed framework transfers informative features from an HR-trained network to an LR-trained network by reducing the distance between them. A cosine similarity measure was employed as a distance metric to effectively align the HR and LR features. This approach differs from conventional knowledge distillation frameworks, which use the L_p distance metrics and offer the advantage of converging well when reducing the distance between features of different resolutions. Our framework achieved a 3% improvement over the previous state-of-the-art method on the AgeDB-30 benchmark without bells and whistles, while maintaining a strong performance on HR images. The effectiveness of cosine similarity as a distance metric was validated through statistical analysis, making our approach a promising solution for real-world applications in which LR images are frequently encountered. The code and pretrained models are publicly available on https://github.com/gist-ailab/feature-similarity-KD.