Although object detection has achieved great success in the field of computer vision in the past few years, the performance of detecting small objects has not yet achieved ideal results. For instance, UAV aerial photography object detection plays an important role in traffic monitoring and other fields, but it faces some great challenges. The objects in aerial images are mainly small objects, the resolution of whom is low and the feature expression ability of whom is very weak. Information will be lost in high-dimensional feature maps, and this information is very important for the classification and positioning of small objects. The most common way to improve small object detection accuracy is to use high-resolution images, but this incurs additional computational costs. To address the above-mentioned problems, this article proposes a new model SINextNet, which uses a new dilated convolution module SINext block. This module is based on depth-separable convolution, and can improve the receptive field of the model. While extracting small object features, it can combine small object features with background information, greatly improving the feature expression ability of small objects. The experimental results indicate that the method proposed in this paper can achieve advanced performance across multiple aerial datasets.
1 June 2024 in Journal of Artificial Intelligence and Soft Computing Research
Self-supervised monocular depth estimation has been widely applied in autonomous driving and automated guided vehicles. It offers the advantages of low cost and extended effective distance compared with alternative methods. However, like automated guided vehicles, devices with limited computing resources struggle to leverage state-of-the-art large model structures. In recent years, researchers have acknowledged this issue and endeavored to reduce model size. Model lightweight techniques aim to decrease the number of parameters while maintaining satisfactory performance. In this paper, to enhance the model’s performance in lightweight scenarios, a novel approach to encompassing three key aspects is proposed: (1) utilizing LeakyReLU to involve more neurons in manifold representation; (2) employing large convolution for improved recognition of edges in lightweight models; (3) applying channel grouping and shuffling to maximize the model efficiency. Experimental results demonstrate that our proposed method achieves satisfactory outcomes on KITTI and Make3D benchmarks while having only 1.6M trainable parameters, representing a reduction of 27% compared with the previous smallest model, Lite-Mono-tiny, in monocular depth estimation.
11 August 2025 in Data Mining and Knowledge Discovery
Most existing low-light image enhancement (LIE) methods are primarily designed for human-vision-friendly image formats, such as sRGB, due to their convenient storage and smaller file sizes. In addition, raw images provide greater detail and a wider dynamic range, which makes them more suitable for LIE tasks. Despite these advantages, raw images, the original format captured by cameras, are larger and less accessible and are hard to use in methods of LIE with mobile devices. In order to leverage both the advantages of sRGB and raw domains while avoiding the direct use of raw images as training data, this paper introduces sRrsR-Net, a novel framework with the image translation process of sRGB–raw–sRGB for LIE task. In our approach, firstly, the RGB-to-iRGB module is designed to convert sRGB images into intermediate RGB feature maps. Then, with these intermediate feature maps, to bridge the domain gap between sRGB and raw pixels, the RAWFormer module is proposed to employ global attention to effectively align features between the two domains to generate reconstructed raw images. For enhancing the raw images and restoring them back to normal-light sRGB, unlike traditional Image Signal Processing (ISP) pipelines, which are often bulky and integrate numerous processing steps, we propose the RRAW-to-sRGB module. This module simplifies the process by focusing only on color correction and white balance, while still delivering competitive results. Extensive experiments on four benchmark datasets referring to both domains demonstrate the effectiveness of our approach.
Advances in deep learning have propelled the evolution of multi-view clustering techniques, which strive to obtain a view-common representation from multi-view datasets. However, the contemporary multi-view clustering community confronts two prominent challenges. One is that view-specific representations lack guarantees to reduce noise introduction, and another is that the fusion process compromises view-specific representations, resulting in the inability to capture efficient information from multi-view data. This may negatively affect the accuracy of the clustering results. In this article, we introduce a novel technique named the “contrastive attentive strategy” to address the above problems. Our approach effectively extracts robust view-specific representations from multi-view data with reduced noise while preserving view completeness. This results in the extraction of consistent representations from multi-view data while preserving the features of view-specific representations. We integrate view-specific encoders, a hybrid attentive module, a fusion module, and deep clustering into a unified framework called AMCFCN. Experimental results on four multi-view datasets demonstrate that our method, AMCFCN, outperforms seven competitive multi-view clustering methods. Our source code is available at https://github.com/xiaohuarun/AMCFCN.
11 April 2025 in Proceedings of the AAAI Conference on Artificial Intelligence
Coordination and joint ability are important topics in representation and reasoning about multi-agent systems. The modal logic JAADL proposed by Liu et al. extends ATL with joint abilities, which enables reasoning about whether a coalition of agents can coordinate and achieve a goal without communication. However, like ATL, strategic abilities in JAADL are defined in terms of combinatorial strategies, which are functions from histories or states to actions. On the other hand, there has been research on reasoning about natural strategic abilities, where a natural strategy is formalized as a sequence of condition-action pairs, making it more human-friendly than combinatorial strategy. In this work, we propose SJAADL, a variation of JAADL where strategic abilities are defined in terms of structured strategies represented with LDL (linear dynamic logic) formulas, with bounded complexity. We use nondeterministic strategies since they are more expressive, natural and succinct than determinstic ones. We present syntax and semantics of SJAADL. We show that model checking SJAADL can be done in time quasi-polynomial with the model size, exponential with the formula size, and with the complexity bound of structured strategies, exponential in the memoryless case and double exponential in the memoryful case. Finally, we introduce the problem of synthesizing norms to achieve joint abilities, and give two algorithms for it.