As one of the important fields of computer vision research, pedestrian attribute recognition has gained increasing attention from domestic and foreign researchers due to its huge potential applications. However, obtaining long-distance pedestrian information in actual scenes poses challenges such as lack of information, incomplete feature extraction, and low attribute recognition accuracy. To address these issues, we propose a multi-scale feature fusion network based on a dual self-attention mechanism. The fusion module merges multi-scale features to enable more complete attribute extraction, while the dual self-attention module focuses the network on important regions. Experimental results on PA-100K, RAP, and PETA datasets achieved mean accuracies of 81.97
Aiming at the problems of complex pre-processing and expensive equipment in chemical detection of graphite ore grade, a graphite ore identification and classification method based on fusing multi-scale image features is proposed. In the feature extraction stage, a deep convolutional neural network and a residual network model based on spatial attention mechanism are constructed to improve the learning ability of local and global features of graphite ore images; in the feature aggregation stage, a global response normalization technique is introduced to achieve more accurate graphite ore grade recognition, and the accuracy of the model reaches 93.401
Human skeleton point data has better environmental adaptability and motion expression ability than RGB video data. Therefore, the action recognition algorithm based on skeletal point data has received more and more attention and research. In recent years, skeletal point action recognition models based on graph convolutional networks (GCN) have demonstrated outstanding performance. However, most GCN-based skeletal action recognition models use three stable spatial configuration partitions, and manually set the connection relationship between each skeletal joint point. Resulting in an inability to better adapt to varying characteristics of different actions. And all channels of the input X features use the same graph convolution kernel, resulting in coupling aggregation. Contrary to the above problems, this paper proposes a new division strategy, which can better extract the feature information of neighbor nodes of nodes in the skeleton graph and adaptively obtain the connection relationship of joint nodes. And introduce Decoupled Graph Convolution (DC-GCN) to each partition to solve the coupled aggregation problem. Experiments on the NTU-RGB+D dataset show that the proposed method can achieve higher action recognition accuracy than most current methods.
For the problem of generating high-quality and diverse images, an image generation method combining residual module, spectral parametric normalization, and self-attention mechanism is proposed to be applied in WGAN networks. The specific improvement of the method is to introduce the residual module into the generator and discriminator networks to better capture the deep image information. The spectral parametric normalization technique is also applied to each convolutional layer of the residual block to improve the stability of the image generation process. The self-attention mechanism is introduced into the generator to enable the network to learn in a targeted manner and generate higher-quality images. The experimental results demonstrate that the combined application of these techniques can effectively solve the challenge of generating image samples, obtain stable and diverse data samples, generate better results than the original WGAN method and DCGAN method, and use the generated data samples as the dataset for expanding the classification experiments, which improves the recognition accuracy of the image classification network to a certain extent.
Light field (LF) camera sensors often face a trade-off between angular resolution and spatial resolution when shooting. High spatial resolution image arrays often result in lower angular resolution, and vice versa. In order to obtain high spatial resolution and at the same time have high angular resolution. In this paper, we propose an improved 4D convolutional neural network (CNN) algorithm for angular super-resolution (SR) to improve the quality of angular SR images. Firstly, to address the problem of low luminance of images captured by LF cameras, this paper uses block threshold square reinforcement (BTSR) for image luminance enhancement. Secondly, to make the reconstructed new viewpoints of higher quality, this paper improves the attention mechanism convolutional block attention module (CBAM). This paper incorporates it into a 4D dense residual network as high dimensional attention module (HDAM). HDAM generates images along two independent dimensions, spatial and channel. The HDAM generates attention maps along two independent dimensions, space and channel, which guide the network to focus on more important features for adaptive feature modification. Finally, this paper modifies the activation function to make the network perform better in the later stages of training and more suitable for LF reconstruction tasks. This paper evaluates the network on many LF data, including real-world scenes and synthetic data. The experimental results show that the improved network algorithm can achieve higher quality LF reconstruction.
