Recently, with the development of deep learning, the field of lightweight object detection has witnessed significant progress. However, mainstream lightweight detectors ignore the extraction of multi-scale semantic information. In addition, these approaches ignore the relationship between deep semantic features and shallow detail features. To relieve above shortcomings, a Pyramid Pooling Enhanced Multi-scale Network(PPMENet) is proposed and an Efficient Pyramid Pooling Block (EPPB) is designed to extract multi-scale deep semantic information,strengthening the feature expression ability of the model. On the other hand, a Cross Semantic Level Interaction Attention Module (CSIAM) is designed to enhance information interaction between features at different semantic levels. Experimental results on the MS COCO 2017 test set show that PPMENet gets 28.0% average precision, only with 2.16×10⁶ model size and 0.97GFLOPs,and achieves inference speed of 218 frame/s. Compared with other methods, PPMENet realizes a good balance between detection accuracy and model execution efficiency.

For the optimal feature subset selection and model parameter optimization in ultra-wideband non-line-of-sight(NLOS) recognition，a new NLOS recognition method based on the cross-validation recursive feature elimination algorithm of Light Gradient Boosting Machine(LightGBM) and Optuna parameter tuning is proposed. First，six important features，including the difference between the first path signal and the total received signal power，and the maximum noise，are selected as the optimal feature subset using the recursive feature elimination and cross-validation algorithm. Then，Optuna is used to optimize the hyperparameters of LightGBM model. Line-of-sight and non-line-of-sight feature data is collected，and the Support Vector Machine，Extreme Gradient Boosting algorithm，and parameter-optimized LightGBM model are trained and tested. The results demonstrate that the selected features exhibit excellent discriminative ability，with the optimized LightGBM model achieving a recognition accuracy of 95.28% .

In order to further improve the accuracy of human action recognition and fully explore the spatiotemporal features of action sequences, a graph convolution action recognition method based on spatiotemporal feature fusion and attention mechanism is proposed. The spatial attention map convolution is used to refine the topology to capture the correlation features of the joints under different motion types,and the time convolution structure is extended by the time domain multi-scale convolution module to capture the multi-scale time features. A multi-level feature fusion module is constructed,which takes the initial feature and the convolution output feature of the time-domain multiscale graph as the module input,and uses a two-branch structure to obtain the global and local channel features respectively. On this basis,a limb attention mechanism is proposed to divide the human topological structure and calculate the attention weights in the channel dimension respectively to enhance the model's ability to pay attention to local action features. The experimental results show that the recognition accuracy is 93.0% and 96.9% in CS and CV evaluation mode of NTU RGB+D data set,and 89.8% and 91.1% in X-Sub and X-Set evaluation mode of NTU RGB+D 120 data set,respectively. The recognition accuracy is higher than that of ST-GCN,CTR-GCN and other models.

In satellite communication systems operating in dynamic interference environments,the quality of channels and the interference power vary. Limited spectrum resources and complex interference environments pose challenges for anti-interference communication decisions, particularly in terms of resource allocation and service demands. Specifically, the challenge lies in efficiently utilizing resources while avoiding interference frequencies and optimizing power. To address this issue,a deep reinforcement learning-based anti-interference algorithm with multiple reward functions is proposed. The algorithm models the interaction between the transmitter,receiver,and interferer as a Markov decision process. By optimizing the reward function associated with the costs of channel and power switching,it introduces mechanisms for both frequency and power switching,analyzes the interference characteristics in the spectrum of adjacent time slots, and integrates the interference signal features collected during the interaction with channel information to train an anti-interference strategy. This strategy enables joint anti-interference decision-making in both the frequency and power domains. Simulation results demonstrate that the algorithm effectively reduces the probability of interference,accelerates convergence,and optimizes the utilization of power resources.

In order to reduce complexity of construction of indoor positioning fingerprint database and improve the positioning accuracy, an indoor fingerprint positioning algorithm based on matrix completion under the 5G ultra-dense network is proposed. In the offline database construction stage,the algorithm first uses the K-nearest Neighbor(KNN) interpolation method to complete the matrix of part of the fingerprint database to construct a complete database. Secondly,the sparse auto-encoder is used to extract the sparse features of the fingerprint database, and the high-dimensional received signal strength indication (RSSI) signal is reduced. In the online fingerprint matching stage,the weighted KNN algorithm is used to estimate the coordinates of the point to be located. After experimental simulation,the average relative error of the algorithm to reconstruct the fingerprint database is 0.31% . Compared with that of the traditional KNN fingerprint matching algorithm,the average error is reduced by 24.41% .

For the problems of limited expression of characteristic information and low classification accuracy in radiation source classification tasks,an individual radiation source recognition method based on multi-resolution feature fusion is proposed. In this method, the individual characteristics of the radiation source are expressed by using three time-frequency spectra with different resolutions obtained through the Short-Time Fourier Transform. Multi-channel convolutional neural networks are constructed using ResNext50 to extract features with different time-frequency resolutions. A multi-channel feature weighted fusion mechanism is introduced into the network,and the features of different channels are fused by feature weighted fusion, combining the feature information from different resolutions. Experiments show that this method improves the ability to express the subtle fingerprint information of the radiation source signal,and compared with that of the feature layer fusion method and the single feature expression method, the recognition accuracy is improved by 2.15% and 6.8% ,respectively.

For the issue of decreased detection performance under low signal-to-noise ratio (SNR) conditions due to insufficient utilization of covariance matrix information in covariance-based eigenvalue algorithms for constructing detection statistics,a novel spectral sensing algorithm based on the ratio of the difference between the maximum and minimum eigenvalues to the harmonic mean of eigenvalues is proposed. This algorithm constructs the detection statistic by incorporating both the extreme eigenvalues and the harmonic mean of eigenvalues from the covariance matrix,thereby more comprehensively exploiting the eigenvalue information within the covariance matrix to enhance the detection capability. Furthermore, a novel approach for calculating the harmonic mean is introduced, leveraging the asymptotic distribution theory of eigenvalues in random matrices. This approach aims to not only improve the accuracy of the decision threshold but also further boost the detection performance. Simulation results demonstrate that the proposed algorithm,without requiring prior knowledge of primary users or channel conditions,achieves a detection probability increase of no less than 10% compared with several classic algorithms at -20 dB SNR.

For the problem that conventional signal analysis methods are prone to cause the“increasing batch”and“missing batch”in multi-functional radar signal sorting in complex electromagnetic environments, a multi-functional radar signal sorting method based on improved complex network community detection is proposed. This method first transforms the signal sequence into the complex network using limited penetration visibility graph. Then, it introduces spatial clustering with density to eliminate spurious pulses. Subsequently, the label propagation algorithm is improved according to the between centrality of nodes,enhancing the stability of community division. Finally,the sub-communities are merged to complete the signal sorting task by density peak clustering. Simulation results show that the proposed method achieves a sorting accuracy of 98.13% for multi-functional radar signals. Moreover,even when the proportion of spurious pulses increases to 35% , the number of sorted batches remains unchanged, effectively alleviating the“increasing batch”and“missing batch”problems.

To address the issue of synthetic aperture radar(SAR) images being susceptible to environmental noise,which leads to a reduction in the signal-to-noise ratio(SNR) at target locations,a method integrating singular value decomposition (SVD) and minimum entropy deconvolution (MED) to enhance the Back Projection(BP ) algorithm is proposed. Initially,the acquired echo signals undergo SVD, and a singular value matrix is obtained. By retaining only the first five singular values and reconstructing the echo signal matrix,an initial noise reduction is achieved. Subsequently,the signals are processed using MED filtering, where filter coefficients are iteratively updated to minimize the entropy of the signal,thereby reducing the kurtosis of the output signal and suppressing noise. A Zero-Phase Filter(ZPF) is then applied to restore any phase delays. Finally, the SAR image is generated using the BP algorithm. Experimental results demonstrate that this method significantly enhances the SNR at target locations,with improvements of 7.9 dB and 9.1 dB for large and small corner reflectors,respectively.

In view of the problems of false detection and missed detection when an unmanned aerial vehicle (UAV) detects targets at different scales,a YOLOv8-FDT UAV algorithm model with a multi-scale fusion mechanism is proposed. First, a dynamic upsampling module is added to the Neck layer of the baseline model to reduce the number of model parameters and improve the real-time performance of the model for target recognition. In addition, in order to enable the entire algorithm model to capture different scale semantic information of the target in the feature fusion stage,adaptive downsampling and depth convolution are integrated to design the feature diffusion pyramid network(FDPN). Finally,experiments on the UAV aerial photography dataset VisDrone2019 show that the mean average precision(mAP) of all categories of the improved model is increased by 6.24% compared with that of the baseline model.