To overcome the deficiencies of low-rank embedding in data reconstruction and noise suppression, and improve the accuracy of its feature recognition,an image feature extraction algorithm is proposed based on orthogonal projection learning. The half-quadratic alternating direction method of multipliers algorithm is designed to solve the orthogonal projection learning model. The model retains the main features of the samples by introducing an orthogonal matrix,the norm constraints makes the extracted features more prominent, and the weighted Schatten p-Norm is used to approximate the optimal solution of the rank. To improve the robustness of the model and make it suitable for supervised scenarios, generalized correntropy is used for data item modeling and classification loss function construction. Experimental results on different scale datasets show that the proposed model has better feature extraction performance than other existing models.
The sensitivity encoding ( SENSE) technique utilizes sensitivity information from multiple receiving coils to reduce scan time. The existing SENSE-based parallel MRI reconstruction methods have problems of artifacts and missing details, which is not conducive to clinical diagnosis. By introducing data-driven adaptive sparse transform learning (TL) into the SENSE algorithm, TL-SENSE algorithm is proposed, that reduce the artifacts and improve the quality of parallel MRI reconstruction. The proposed algorithm employs the alternating direction method of multipliers (ADMM) to solve the target optimization problem. And The proposed algorithm comprises three steps: transform updating, hard threshold denoising and image updating. The simulation results show that the proposed algorithm performs well in image denoising and restoration and preserves the texture details and edge information. It also achieves higher consistency between the reconstructed image and the original image. For the selected 48 sets of data, the average signal noise ratio of TL-SENSE increased by 4.62 dB, 1.91 dB, 1.30 dB and 0.89 dB compared with that of SENSE, L1-SENSE, TV-SENSE, and LpTV-SENSE, respectively.
A YOLOv3-based multi-target detection method is proposed to address the problem of missed and false detection caused by the small percentage of target pixels and mutual occlusion in traffic scenes. The method implants a spatial pyramid pooling module in the YOLOv3 network structure to enhance feature representation, and a multi-scale feature fusion mechanism is proposed to obtain both spatial information and semantic information. The semantic information of the target to be detected is refined by extending the prediction branch of the prediction layer. In addition, the improved K-means + + clustering algorithm is used to extract the initial center of the priori box and improve the matching degree between the prediction anchor box and the target to be detected. Meanwhile, a flexible non-maximum suppression algorithm is applied to adjust the confidence score flexibly. The experimental results based on the hybrid data set show that the proposed method improves the detection accuracy effectively.
In the case of limited measurement information, Kalman filter (KF) is difficult to deal with unmanned aerial vehicle tracking by using a single motion model. To solve this problem, a novel long short-term memory(LSTM)-KF algorithm combining LSTM and KF algorithm is proposed. First, LSTM is used to predict the average and instantaneous velocity of the target so that the problem of poor generalization ability of nonparametric model can be solved in position prediction task. Then, the prediction limitation of KF algorithm using motion model is analyzed, and the method of using LSTM prediction results to modify the prediction results of motion model is proposed to reduce the prediction error. The revised prediction results are combined with the measurement data to realize the state estimation of the target. Finally, the proposed algorithm is verified on the generated trajectory. The simulation results show that LSTM-KF algorithm has higher tracking accuracy and stronger robustness than the existing models.