To tackle the challenges of inadequate feature representation, detail degradation, and imbalanced class distribution in road extraction from remote sensing imagery, this research introduces a road extraction model from remote sensing images based on attention and adaptive feature fusion Unet (AAF-Unet). The model uses VGG16 as the backbone network. Firstly, the cross-spatial efficient multi-scale attention (CEMA) module is added at the skip connection of the Unet network, effectively integrating global contextual information and enhancing the model's feature perception ability for complex backgrounds and occluded areas. Secondly, add the adaptive convolution mix (ACmix) module at the bridge between the encoder and the decoder. By combining dynamic convolution with self-attention mechanisms, this module enables efficient integration of local detail features and global context information, significantly improving the model's performance in edge and detail recovery. In the loss function section a composite loss function combining Dice loss and Focal loss is proposed, optimizing the model's performance in segmentation tasks involving small targets and edge-blurred regions. Lastly, through the experimental verification on different dataset, the mIoU index of AAF-Unet model reached 82.01% on DeepGlobe dataset and 79.36% on the CHN6-CUG dataset. Compared with Unet, Deeplabv3+ and Hrnet models, the extraction results were increased respectively on two dataset. Good segmentation accuracy was achieved while ensuring the generalization ability of the model.

Aiming at the low accuracy of traditional low-resolution face recognition, a low-resolution face recognition network InGLRNet (inception GAN low resolution Net) based on super-resolution reconstruction and common feature subspace is proposed. The InGLRNet network adopts a generative adversarial network structure. First, the generation network is composed of the super-resolution InSRNet network and the feature extraction network. The InSRNet network decomposes the 3×3 convolution kernel in the Inception structure into 1×3, 3×1 and 1×1 convolution kernels, and at the same time increases the bypass direct connection of residual network, which can alleviate the problem of gradient disappearance. Secondly, using the common feature subspace method, the distance between the transformed low-resolution sample image and the high-resolution reference image in the common space is used as the objective function for training a deep convolutional neural network, and the loss function is used to match the high and low features of high-resolution images to achieve accurate feature recognition of faces. Finally, the InGLRNet is compared with the four classic low-resolution face recognition methods, CLPMs, MDS, Deep-Face and Face-Net. The experimental results show that the constructed network has a significant improvement in face recognition performance. It is better than the other 4 methods at different low resolution.

In order to achieve remote sensing image small object detection task, there are problems such as poor feature extraction and fusion due to the blurring of target detail texture information, and leakage of small targets, a remote sensing objects detection algorithm that leverages multiscale feature enhancement and interactive fusion is proposed. Firstly, the multiscale feature enhancement (MFE) module with cross-layer and multi-branch connection structure is used to enrich and enhance the texture feature information obtained from different gradients by using the Split shunt operation, and at the same time, the lightweight feature phantom module Ghost is introduced to perform the linear transformation of the channel, generating more effective feature detail information flow to enhance attention to local detail feature information in the image. Secondly, the feature interaction fusion (FIF) module is constructed, which introduces a multi-branch serial parallel convolution block and an adaptive mechanism pooling block to interact with the channel semantic information and spatial transformation of the input features, capture the global context information, and accurately locate the key position of the small targets. information, enhance the correlation between feature information, and achieve multi-dimensional interactive fusion of fine-grained features. The proposed algorithm is validated with DIOR, which is a remote sensing dataset. The optimized network model achieves a mean accuracy precision of 87.6%, which is higher than the other seven excellent algorithms including NPMMR-Det, YOLOv7, and YOLOv5. The improved small target detection algorithm for remote sensing images achieves better detection accuracy.

A real-time UAV image segmentation algorithm with enhanced contextual feature interaction is proposed to address the problem of target omission and incompleteness in segmentation results due to the lack of global information interaction in lightweight algorithms for semantic segmentation tasks of UAV images. The approach uses a two-branch structure. To encode the channel and spatial information, global average pooling in various directions was used. This preserves the correct position information and increases the attention to the local detail information in the image. Secondly, a global perceptual extraction module was constructed by using the position-aware circular convolution and spatial weighting, which achieves the global contextual information capture; Finally, the weighting operation is applied to the features of different scales for the fusion, which reduces the information loss in the fusion process and the computation of the algorithm. The UAVid and AeroScapes datasets are used to validate the algorithm. The results indicate that the mean intersection over union (mIoU) achieved 66.5% and 63.0%, respectively, marking a 2.6% and 2.2% improvement over BiSeNet V2. The segmentation speeds reached 79.9 and 71.4 frames per second, respectively, showing an increase of 8.3 and 6.9 frames per second compared to BiSeNet V2. This method ensures real-time segmentation speed while delivering satisfactory segmentation accuracy.