Early Behavioral Perception Of Alzheimer's Disease Based On Deep Learnin

Alzheimer’s disease is a neurodegenerative disease,and its main clinical manifestations are the loss of emotional management ability,the decline of memory and the deterioration of daily activities.Alzheimer’s disease is a serious threat to human health and quality of life,but it is a pity that no effective drugs or treatments can reverse the progress of Alzheimer’s disease so far.If clues can be found before the critical state of "no disease",timely diagnosis and intervention treatment can be implemented,so the study of early behavioral perception of Alzheimer’s disease is of great significance.In this paper,the key frame recommendation method based on multi-feature intersection and fusion model is studied,and based on this method,a video data processing algorithm is proposed to deal with massive daily video data.A data set of Alzheimer’s disease premonitory action based on daily activity video data is constructed.A recognition method of Alzheimer’s disease premonitory action based on spatiotemporal two-stream network is designed.The main research contents of this paper are as follows:(1)A key frame recommendation method based on multi-feature intersection and fusion model is proposed(FFMKR).To construct a combination of 15 features for each frame of the video,we designed a content feature acquisition algorithm based on the joint extraction of multiple image descriptors and a user feature acquisition algorithm based on inter-frame distance and collaborative filtering;to acquire key frames,we proposed a recommendation model based on multi-feature intersection and fusion.The ablation experimental results show that the ablation of each sub-module of the recommendation model decreases the evaluation index by 1.32% on average,and the ablation of 15 features obtained by the two feature acquisition algorithms reduces the evaluation index by 0.47% on average.The comparison experiments show that the evaluation index of FFMKR is higher than the comparison algorithm by 2.01%,1.66% and 0.34% on average on the HMDB-51,UCF-101 and VSUMM datasets,while the average error of key frame localization is 5.61 frames and6.48 frames on the HMDB-51 and UCF-101 datasets,respectively.(2)An Alzheimer’s disease premonitory action dataset based on video data of daily activities is constructed(ADP).We selected twenty types of Alzheimer’s disease actions which can reflect the decline of human function in daily activities.And 2301 videos were collected with an average of 138 frames and a total of 317,553 frames.To eliminate the redundancy of the original acquisition data,we constructed a FFMKR-based video data processing algorithm and the ADP dataset achieves an average compression rate of 26.8%for 37 frames.(3)A spatiotemporal two-stream network-based approach for premonitory action recognition in Alzheimer’s disease is proposed(STADP).We constructed a spatial feature extraction module based on the 3D convolutional model and temporal feature extraction module based on the Transformer model,and the spatiotemporal two-stream network is a fusion of the above two feature extraction module.To find the best spatiotemporal feature fusion coefficient α,we performed experiments on the ADP dataset.When the optimal α is0.35,the action recognition accuracy of STADP is 83.21% with a variance of 2.949%.The comparison experiments show that STADP is higher than the comparison model in accuracy,precision,recall and F1 value with 5.04%,1.6%,7.99% and 9.07% on average,respectively.