Aiming at the problems of slow detection speed, low precision, missed detection and false detection of current forest pest detection methods, a forest pest detection method based on attention model and lightweight YOLOv4 was proposed. Firstly, a dataset was constructed and preprocessed by using geometric transformation, random color dithering and mosaic data augmentation techniques. Secondly, the backbone network of YOLOv4 was replaced with a lightweight network MobileNetV3, and the Convolutional Block Attention Module (CBAM) was added to the improved Path Aggregation Network (PANet) to build the improved lightweight YOLOv4 network. Thirdly, Focal Loss was introduced to optimize the loss function of the YOLOv4 network model. Finally, the preprocessed dataset was input into the improved network model, and the detection results containing pest species and location information were output. Experimental results show that all the improvements of the network contribute to the performance improvement of the model; compared with the original YOLOv4 model, the proposed model has faster detection speed and higher detection mean Average Precision (mAP), and effectively solves the problem of missed detection and false detection. The proposed new model is superior to the existing mainstream network models and can meet the precision and speed requirements of real?time detection of forest pests.
Aiming at the problems of large model parameters, high computational complexity and low accuracy of traditional violence detection methods, a method of violence detection in video based on temporal attention mechanism and EfficientNet was proposed. Firstly, the foreground image obtained by preprocessing the dataset was input to the network model to extract the video features, meanwhile, the frame-level spatial features of violence were extracted by using the lightweight EfficientNet, and the global spatial-temporal features of the video sequence were further extracted by using the Convolutional Long Short-Term Memory (ConvLSTM) network. Then, combined with temporal attention mechanism, the video-level feature representations were obtained. Finally, the video-level feature representations were mapped to the classification space, and the Softmax classifier was used to classify the video violence and output the detection results, realizing the violence detection of video. Experimental results show that the proposed method can decrease the number of model parameters, reduce the computational complexity, increase the accuracy of violence detection and improve the comprehensive performance of the model with limited resources.
Since there is no perfect theoretical basis for the selection of kernel function in single kernel network models, and the network node size of Four-layer Neural Network based on Randomly Feature Mapping (FRFMNN) is excessively large, a Four-layer Multiple Kernel Neural Network based on Randomly Feature Mapping (MK-FRFMNN) algorithm was proposed. Firstly, the original input features were transformed into randomly mapped features by a specific random mapping algorithm. Then, multiple basic kernel matrices were generated through different random kernel mappings. Finally, the synthetic kernel matrix formed by basic kernel matrices was linked to the output layer through the output weights. Since the weights of random mapping of original features were randomly generated according to the random continuous sampling probability distribution randomly, without the need of updates of the weights, and the weights of the output layer were quickly solved by the ridge regression pseudo inverse algorithm, thus avoiding the time-consuming training process of the repeated iterations. Different random weight matrices were introduced into the basic kernel mapping of MK-FRFMNN. the generated synthetic kernel matrix was able to not only synthesize the advantages of various kernel functions, but also integrate the characteristics of various random distribution functions, to obtain better feature selection and expression effect in the new feature space. Theoretical and experimental analyses show that, compared with the single kernel models such as Broad Learning System (BLS) and FRMFNN, MK-FRMFNN model has the node size reduced by about 2/3 with stable classification performance; compared with mainstream multiple kernel models, MK-FRMFNN model can learn large sample datasets, and has better performance in classification.
The aim of domain adaptation is to use the knowledge in a labeled (source) domain to improve the model classification performance of an unlabeled (target) domain, and this method has achieved good results. However, in the open realistic scenes, the target domain usually contains unknown classes that are not observed in the source domain, which is called open set domain adaptation problem. For such challenging scene setting, the traditional domain adaptation algorithm is powerless. Therefore, an open set fuzzy domain adaptation algorithm via progressive separation was proposed. Firstly, based on the open set fuzzy domain adaptation algorithm with membership degree introduced, the method of separating the known and unknown class samples in the target domain step by step was explored. Then, only known classes separated from the target domain were aligned with the source domain, so as to reduce the distribution difference between the two domains and perform the fuzzy adaptation. The negative transfer effect caused by the mismatch between unknown and known classes was reduced well by the proposed algorithm. Six domain transformation experimental results on Office dataset show that, the accuracy of the proposed algorithm has the significant improvement in image classification compared with the traditional domain adaptation algorithm, and verify that the proposed algorithm can gradually enhance the accuracy and robustness of the domain adaptation classification model.
In order to improve the robustness of visual tracking algorithm when the target appearance changes rapidly, a particle filter tracking algorithm based on adaptive subspace learning was presented in this paper. In the particle filter framework, this paper established a state decision mechanism, chose the appropriate learning method by combining the verdict and the characteristics of the Principal Component Analysis (PCA) subspace and orthogonal subspace. It not only can accurately, stably learn target in low dimensional subspace, but also can quickly learn the change trend of the target appearance. For the occlusion problem, robust estimation techniques were added to avoid the impact of the target state estimation. The experimental results show that the algorithm has strong robustness in the case of illumination change, posture change, and occlusion.