Yuanmeng Song

To achieve different compression rates for images while preserving important regions, a semantic network-based deep residual variational auto-encoder is introduced in this paper. The network is divided into two components, a semantic analysis network and an image compression network. The former evaluates the importance of image pixels and accurately locates important semantic regions and key information in the image. According to the semantic importance of different regions, the encoding strategy is dynamically adjusted. The latter utilizes a deep residual variational autoencoder to efficiently encode and decode images, while combining Lagrange multipliers to adjust the model flexibly and the weights of the bitrate. With this model, multiple compression rates have been implemented. Quality of reconstruction achieves better performance with compression at different rates. Finally, a semantic loss function is proposed to replace traditional compression loss functions. Extensive experiments conducts on several datasets, including the Kodak, CLIC and Tecnick TESTIMAGES, which were public datasets. Results demonstrate that our method can effectively improve the quality of image reconstruction at various compression rates. Compared with traditional methods, the peak signal-to-noise ratio is increased by an average of 2dB at the same bitrate, and the structural similarity index is the most close to 0.998. The subjective visual quality is better, especially when processing complex scene images, which can better preserve the details and textures of key objects. This approach can effectively avoid common distortions such as block effects and blurring in traditional methods.

To detect and recognize small-size and submerged complex background targets in infrared images, we combine a dynamic receptive field fusion strategy and a multi-scale feature fusion mechanism to improve the detection performance of small targets significantly. The space-to-depth convolution module is introduced as a downsampling layer in the backbone first and achieves the same sampling effect. More detailed information is retained at the same time. Thus, the model’s detection capability for small targets has been enhanced. Then, the pyramid level 2 feature map with minimum receptive field and maximum resolution is added to the neck, which reduces the loss of positional information during feature sampling. Furthermore, x-small detection heads are added, the understanding of the overall characteristics and structure of the target is enhanced much more, and the representation and localization of small targets have been improved. Finally, the cross-entropy loss function in the original network model is replaced by an adaptive threshold focal loss function, forcing the model to allocate more attention to target features. The above methods are based on a public tool, the eighth version of You Only Look Once (YOLO) improved, it is named SPT–YOLO (SPDConv + P2 + Adaptive Threshold + YOLOV8s) in this paper. Some experiments on datasets such as infrared small object detection (IR-SOD) and infrared small target detection 1K(IRSTD-1K), etc. have been executed to verify the proposed algorithm; and the mean average precision of 94.0% and 69% under the condition of threshold at 0.5 and over a range from 0.5 to 0.95 is obtained, respectively. The results show that the proposed method achieves the best performance of infrared small target detection compared to existing methods.

In order to provide better geographic information services on time, a detection technique on the variation of remote sensing image has been investigated in this paper. Based on multiscale feature multiplex, a consistent regularization method together with a semi-supervised strategy has been proposed, which can solve the problems of unclear change edges and the issue of missing detection of very small targets. By constructing a learning training model of detection network, the performances have been enhanced with limited labeled data and effectively handle change detection tasks in complex scenes. The network consists of two parts. One part is the supervised multiscale feature fusion and multiplex, which enhances feature expression capabilities using channel and spatial attention mechanisms. We extract variation features through complex multiscale feature fusion and reuse strategies to alleviate unclear change edges and the omission of very small targets. The other part is the consistency regularization method for semi-supervised learning model, which adopts an alternating training among models. The prediction differences have been minimized, then updated model parameters. This constructed model of detection network has been verified at different labeling ratios on the datasets such as Wuhan university, China, building change detection and google earth imagery Guangzhou, China, change detection, etc. The experimental results show that this proposed technique demonstrates the best overall performance in complex situations with small dataset labeling ratios among existing detection methods.

To detect and recognize small size and submerged complex background target in infrared images, we combine a dynamic receptive field fusion strategy and a multi-scale feature fusion mechanism to improve the detection performance of small targets significantly. The space-to-depth convolution module is introduced as a downsampling layer in the backbone firstly, and achieved the same sampling effect. More detailed information retained at the same time. Thus the model's detection capability for small targets has been enhanced. Then the pyramid level 2 feature map with minimum receptive field and maximum resolution is added to the neck, which reduces the loss of positional information during feature sampled. Furthermore, x-small detection head are added, The understanding of the overall characteristics and structure of the target is enhanced much more, and the representation and localization of small targets has been improved. Finally, the cross-entropy loss function in the original network model is replaced by an adaptive threshold focal loss function, forcing the model to allocate more attention to target features. Above methods base on a public tool, the 8th version of You Only Look Once (YOLO) improved, it is named SPT-YOLO in this paper. Some experiments on datasets such as infrared small object detection (IR-SOT) and infrared small target detection 1K (IRSTD-1K), etc, have been executed to verify the proposed algorithm; and the mean average precision of 94.0% and 69% under condition of threshold at 0.5 and over a range from 0.5 to 0.95 is obtained, respectively. The results show that the proposed method achieves the best performance of infrared small target detection compared to existing methods.

Remote sensing image change detection is a key application technology in remote sensing. In recent years, the development of deep learning has brought new opportunities for remote sensing image change detection. However, to achieve high change detection accuracy, large-scale, good-quality labeled data is required to train the model, which is often difficult to obtain and costly. This paper proposes a semi-supervised remote sensing image change detection network based on consistent regularization for multi-scale feature fusion and reuse (SCMFM-CDNet), which can enhance change detection performance with limited labeled data. Specifically, the SCMFM-CDNet consists of two parts, the supervised stage multi-scale feature fusion and reuse network (MFM-CDNet) and the consistency regularization method of semi-supervised learning. The MFM-CDNet network uses channel attention mechanism and spatial attention mechanism to enhance the expression ability of features, and gradually completes the extraction and detection of change features through multi-scale feature fusion and reuse strategy. The consistency regularization method of semi-supervised learning adopts the teacher model and student model for alternate training, updating model parameters by minimizing the differences in their predictions, thus achieving semi-supervised learning. This paper conducted extensive experiments on three public datasets, LEVIR-CD, WHU-CD, and GoogleGZ-CD. Under the three labeling data ratios of 5%, 10%, and 20%, the F1 scores of the SCMFM-CDNet network on the LEVIR-CD dataset were 88.62, 90.78, and 90.63, respectively. On the WHU-CD dataset, the F1 scores were 88.40, 90.17, and 90.83, respectively. On the GoogleGZ-CD dataset, the F1 scores were 81.44, 82.36, and 85.40, respectively. The results indicate that the proposed network outperforms existing semi-supervised change detection methods in terms of comprehensive performance. In addition, this paper also conducted ablation experiments to evaluate the contribution of the global context module (GCM) and feature fusion reuse module (FFM) in the MFM-CDNet network. The experimental results show that these two modules are crucial for improving network performance.