The widespread application of mobile robots holds significant importance for advancing social intelligence. However, as the complexity of the environment increases, existing Obstacle Avoidance Path Planning (OAPP) methods tend to fall into local optimal paths, compromising reliability and practicality. Therefore, based on the Spider-Wasp Optimizer (SWO), this paper proposes an improved OAPP method called the LMBSWO to address these challenges. Firstly, the learning strategy is introduced to enhance the diversity of the algorithm population, thereby improving its global optimization performance. Secondly, the dual-median-point guidance strategy is incorporated to enhance the algorithm’s exploitation capability and increase its path searchability. Lastly, a better guidance strategy is introduced to enhance the algorithm’s ability to escape local optimal paths. Subsequently, the LMBSWO is employed for OAPP in five different map environments. The experimental results show that the LMBSWO achieves an advantage in collision-free path length, with 100% probability, across five maps of different complexity, while obtaining 80% fault tolerance across different maps, compared to nine existing novel OAPP methods with efficient performance. The LMBSWO ranks first in the trade-off between planning time and path length. With these results, the LMBSWO can be considered as a robust OAPP method with efficient solving performance, along with high robustness.

In this paper, we propose a deep neural structure for detecting image copy and paste tampering, including basic copy-move types and copy-move types that overlay post-processing operations. Based on multi-scale Swin Transformer and dense upsampling convolution, we can effectively detect tampered images in small areas. By introducing multi-scale Swin Transformer feature extraction network and integrating global features and local features, the network can adapt to various shapes and sizes of tampered areas, especially in small areas. The effect of tampering is significantly improved. At the same time, the dense up-sampling convolution is used, and the multi-channel filter is used to amplify the down-sampling feature map to restore the input size. The experimental results show that on the public image tamper detection benchmark, this method has significantly improved compared with the comparison method. Compared with BusterNet2019, the accuracy rate, recall rate and value have increased by 16.74 percentage points, 16.48 percentage points and 16.68 percentage points respectively, and the effect has been improved more significantly in small area tampering.

Accurately predicting rainfall runoff is very beneficial for flood forecasting, water resource management, and planning. Due to the nonlinear time series nature of this problem, it remains a challenging problem in hydrological information processing. However, some methods fail to provide better forecasts for the rainy and dry seasons in certain regions. To address this issue, this paper proposes a rainfall-runoff prediction model called P-former, which is based on two kinds of positional encodings: Feature Positional Encoding (FPE) and Temporal Positional Encoding (TPE). FPE and TPE capture the correlation between meteorological data characteristics and the periodicity of time information, thereby improving the model's ability to predict future rainy and dry seasons, sensing the range of runoff changes, and ultimately improve its runoff prediction ability. Experimental results demonstrate that the P-former model decreases the Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) scores of the single-day average runoff forecasts in the rainy season by 48%-53% and 11%-19%, respectively, in terms of both performance and accuracy.

As an important research direction in the field of natural language processing, multimodal sentiment analysis is an essential part of semantic analysis, natural language understanding and other downstream tasks. Aiming at the problem that the feature differences between different modes in the existing methods make it difficult to fuse and utilize the cross-feature information, this paper proposes a cross-modal feature fusion model based on cross-attention mechanism and self-attention mechanism (multi-modal feature fusion, MMFF). The method uses separated feature extraction networks to encode different modalities and obtain unimodal features. In terms of cross-modal feature fusion, the cross-attention mechanism and the self-attention mechanism are used to obtain inter-modal interaction features and contextual features, and performs cross-modal fusion of different unimodal features to obtain the weights and correlations of the different semantic parts of the multiple feature sequences, so as to enhance the information interaction between different modalities. Multiple experiments on customized microblog Traffic opinion sentiment dataset and CH-SIMS dataset show that the proposed method can effectively utilize multimodal feature, and the experimental results are superior to most baseline multimodal sentiment analysis algorithms, which can effectively improve the performance of multimodal sentiment analysis task.

With the wide application of head-mounted display devices in virtual reality display, the information perception and visual fidelity of VR have been further studied. The key and main advantage of VR display is the ability to seamlessly integrate digital and physical space, which is a new media medium that can provide users with unprecedented immersion experience. However, to save computing resources, the existing virtual reality display rendering methods often choose the foveated rendering. The common foveated rendering method uses the characteristics of human vision to save a lot of computing resources, but almost completely ignores the information perception ability of peripheral vision and visual experience it can bring. This paper proposes a method based on peripheral encoding, which improves the quality of images rendered by the foveated rendering method.