Small-object detection is a challenging task in computer vision due to the limited training samples and low-quality images. Transfer learning, which transfers the knowledge learned from a large dataset to a small dataset, is a popular method for improving performance on limited data. However, we empirically find that due to the dataset discrepancy, directly transferring the model trained on a general object dataset to small-object datasets obtains inferior performance. In this paper, we propose TranSDet, a novel approach for effective transfer learning for small-object detection. Our method adapts a model trained on a general dataset to a small-object-friendly model by augmenting the training images with diverse smaller resolutions. A dynamic resolution adaptation scheme is employed to ensure consistent performance on various sizes of objects using meta-learning. Additionally, the proposed method introduces two network components, an FPN with shifted feature aggregation and an anchor relation module, which are compatible with transfer learning and effectively improve small-object detection performance. Extensive experiments on the TT100K, BUUISE-MO-Lite, and COCO datasets demonstrate that TranSDet achieves significant improvements compared to existing methods. For example, on the TT100K dataset, TranSDet outperforms the state-of-the-art method by 8.0% in terms of the mean average precision (mAP) for small-object detection. On the BUUISE-MO-Lite dataset, TranSDet improves the detection accuracy of RetinaNet and YOLOv3 by 32.2% and 12.8%, respectively.

It is known that modes in axially uniform waveguides exhibit backward-propagation characteristics for which group and phase velocities have opposite signs. For elastic plates, group velocities of backward Lamb waves depend only on Poisson's ratio. This paper explores ways to achieve a large group velocity of a backward mode in hollow cylinders by changing the outer to inner radius ratio, in order that such a mode with strong backward-propagation characteristics may be used in acoustic logging tools. Dispersion spectra of guided waves in hollow cylinders of varying radii are numerically simulated to explore the existence of backward modes and to choose the clearly visible backward modes with high group velocities. Analyses of group velocity characteristics show that only a small number of low order backward modes are suitable for practical use, and the radius ratio to reach the highest group velocity corresponds to the accidental degeneracy of neighboring pure transverse and compressional modes at the wavenumber k = 0. It is also shown that large group velocities of backward waves are achievable in hollow cylinders made of commonly encountered materials, which may bring cost benefits when using acoustic devices which take advantage of backward-propagation effects.

The task of Audio-Visual Event (AVE) localization involves the temporal localization of both audible and visible events captured by camera sensors. However, the audio noise and visual background inherent in these sensors can obscure the distinction between event categories, or the inconsistency between audio and visual information may lead to incorrect predictions of event boundaries. To address these challenges, we propose a cross-modal feature enhancement network designed to enhance modality-specific audio and visual features as well as overall audio-visual event features. Specifically, our modality-specific feature enhancement approach employs cross-modal attentional mechanisms that allow the visual and audio modalities to focus on event-related features within the opposite modality. This approach also establishes intra-modal dependencies within each modality, enhancing the temporal relevance of event-related audio and visual features. Additionally, our audio-visual event feature enhancement scheme develops inter-modal correspondences through audio-visual interaction fusion, aligning the feature representations of audio-visual events temporally. To further mitigate the interference of visual background noise, we have developed an audio-visual channel-spatial fusion module that leverages audio-guided visual attention to improve the visual representation of the sounding target within the visual scene. Experiments conducted on the AVE dataset demonstrate that our method surpasses state-of-the-art methods in terms of AVE localization accuracy in both supervised and weakly supervised settings.

The task of coal-rock cutting event localization is to locate the start and end time of both visible and audible cutting events through video sensors, as well as identify the coal-rock cutting categories. The development of video sensors equipped with functionalities to localize coal-rock cutting events is crucial for the intelligent development of coal mines. However, challenges such as coal dust and equipment noise compromise the discriminability of visual and audio features of cutting events, consequently diminishing the accuracy of event localization. To mitigate these challenges, we propose the Cross-Modal Spatial-Temporal-Semantic Adaptive Network (CM-STSAN). By introducing the Cross-Modal Spatial-Temporal-Semantic Adapter (CM-STSA) into the frozen pre-trained Swin Transformer (Swin-T) model, which enhances the representation of visual and audio features, thereby improving the accuracy of coal-rock cutting event localization. Specifically, initially, a modality-specific temporal relation module is employed to capture the temporal dependencies between the cutting sound source and the cutting action. Subsequently, a spatial audio-visual interaction fusion module is designed to utilize the clues of the opposite modality to achieve cross-modal feature alignment. Lastly, a semantic audio-visual interaction fusion module is devised to use the shared audio-visual semantic interaction to supports accurate inference of cutting semantic. In the latter two steps, an audio-visual attention interaction module dynamically adjusts the weighting of cross-modal features via an attention gating mechanism, effectively suppressing interference from coal dust and background noise. Our methodology verifies the accuracy of coal-rock cutting events localization on the self-bulit Audio-Visual Cutting Event (AVCE) dataset. Furthermore, validation on the publicly available Audio-Visual Event (AVE) dataset demonstrates that our CM-STSAN achieves state-of-the-art results in AVE localization tasks.