Diagnostic results can be radically influenced by the quality of 2D ovarian-tumor ultrasound images. However, clinically processed 2D ovarian-tumor ultrasound images contain many artificially recognized symbols, such as fingers, crosses, dashed lines, and letters which assist artificial intelligence (AI) in image recognition. These symbols are widely distributed within the lesion’s boundary, which can also affect the useful feature-extraction-utilizing networks and thus decrease the accuracy of lesion classification and segmentation. Image inpainting techniques are used for noise and object elimination from images. To solve this problem, we observed the MMOTU dataset and built a 2D ovarian-tumor ultrasound image inpainting dataset by finely annotating the various symbols in the images. A novel framework called mask-guided generative adversarial network (MGGAN) is presented in this paper for 2D ovarian-tumor ultrasound images to remove various symbols from the images. The MGGAN performs to a high standard in corrupted regions by using an attention mechanism in the generator to pay more attention to valid information and ignore symbol information, making lesion boundaries more realistic. Moreover, fast Fourier convolutions (FFCs) and residual networks are used to increase the global field of perception; thus, our model can be applied to high-resolution ultrasound images. The greatest benefit of this algorithm is that it achieves pixel-level inpainting of distorted regions without clean images. Compared with other models, our model achieveed better results with only one stage in terms of objective and subjective evaluations. Our model obtained the best results for 256 × 256 and 512 × 512 resolutions. At a resolution of 256 × 256, our model achieved 0.9246 for SSIM, 22.66 for FID, and 0.07806 for LPIPS. At a resolution of 512 × 512, our model achieved 0.9208 for SSIM, 25.52 for FID, and 0.08300 for LPIPS. Our method can considerably improve the accuracy of computerized ovarian tumor diagnosis. The segmentation accuracy was improved from 71.51% to 76.06% for the Unet model and from 61.13% to 66.65% for the PSPnet model in clean images.
Recent CNNs (convolutional neural networks) have become more and more compact. The elegant structure design highly improves the performance of CNNs. With the development of knowledge distillation technique, the performance of CNNs gets further improved. However, existing knowledge distillation guided methods either rely on offline pretrained high-quality large teacher models or online heavy training burden. To solve the above problems, we propose a feature-sharing and weightsharing based ensemble network (training framework) guided by knowledge distillation (EKD-FWSNet) to make baseline models stronger in terms of representation ability with less training computation and memory cost involved. Specifically, motivated by getting rid of the dependence of offline pretrained teacher model, we design an end-to-end online training scheme to optimize EKD-FWSNet. Motivated by decreasing the online training burden, we only introduce one auxiliary classmate branch to construct multiple forward branches, which will then be integrated as ensemble teacher to guide baseline model. Compared to previous online ensemble training frameworks, EKD-FWSNet can provide diverse output predictions without relying on increasing auxiliary classmate branches. Motivated by maximizing the optimization power of EKD-FWSNet, we exploit the representation potential of weight-sharing blocks and design efficient knowledge distillation mechanism in EKD-FWSNet. Extensive comparison experiments and visualization analysis on benchmark datasets (CIFAR-10/100, tiny-ImageNet, CUB-200 and ImageNet) show that self-learned EKD-FWSNet can boost the performance of baseline models by large margin, which has obvious superiority compared to previous related methods. Extensive analysis also proves the interpretability of EKD-FWSNet. Our code is available at https://github.com/cv516Buaa/EKD-FWSNet.
The extraction of a target speaker from mixtures of different speakers has attracted extensive amounts of attention and research. Previous studies have proposed several methods, such as SpeakerBeam, to tackle this speech extraction problem using clean speech from the target speaker to provide information. However, clean speech cannot be obtained immediately in most cases. In this study, we addressed this problem by extracting features from the electroglottographs (EGGs) of target speakers. An EGG is a laryngeal function detection technology that can detect the impedance and condition of vocal cords. Since EGGs have excellent anti-noise performance due to the collection method, they can be obtained in rather noisy environments. In order to obtain clean speech from target speakers out of the mixtures of different speakers, we utilized deep learning methods and used EGG signals as additional information to extract target speaker. In this way, we could extract target speaker from mixtures of different speakers without needing clean speech from the target speakers. According to the characteristics of the EGG signals, we developed an EGG_auxiliary network to train a speaker extraction model under the assumption that EGG signals carry information about speech signals. Additionally, we took the correlations between EGGs and speech signals in silent and unvoiced segments into consideration to develop a new network involving EGG preprocessing. We achieved improvements in the scale invariant signal-to-distortion ratio improvement (SISDRi) of 0.89 dB on the Chinese Dual-Mode Emotional Speech Database (CDESD) and 1.41 dB on the EMO-DB dataset. In addition, our methods solved the problem of poor performance with target speakers of the same gender and the different between the same gender situation and the problem of greatly reduced precision under the low SNR circumstances.
With the rapid development of artificial intelligence and deep learning technology, automatic speech recognition technology is experiencing new vitality [...]
Speech emotion recognition (SER) is an important component of emotion computation and signal processing. Recently, many works have applied abundant acoustic features and complex model architectures to enhance the model’s performance, but these works sacrifice the portability of the model. To address this problem, we propose a model utilizing only the fundamental frequency from electroglottograph (EGG) signals. EGG signals are a sort of physiological signal that can directly reflect the movement of the vocal cord. Under the assumption that different acoustic features share similar representations in the internal emotional state, we propose cross-modal emotion distillation (CMED) to train the EGG-based SER model by transferring robust speech emotion representations from the log-Mel-spectrogram-based model. Utilizing the cross-modal emotion distillation, we achieve an increase of recognition accuracy from 58.98% to 66.80% on the S70 subset of the Chinese Dual-mode Emotional Speech Database (CDESD 7-classes) and 32.29% to 42.71% on the EMO-DB (7-classes) dataset, which shows that our proposed method achieves a comparable result with the human subjective experiment and realizes a trade-off between model complexity and performance.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.