Sparse Autoencoder Dimension Reduction And Object Detection Assist

With the exponential growth of big data,people must process and analyze more and more data.Various observation methods can make the observed objects present rich details.Since data processing requires the most e?ective,comprehensive,and concise extraction of target features,dimensionality reduction technology faces huge challenges:(1)The problem of dimensionality disaster caused by too many dimensional details of data.(2)Due to data Labels may be noisy,and dimensionality reduction methods based on data labels may cause poor performance when performing other machine learning tasks.(3)For data of di?erent types,dimensions,and application fields,di?erent targeted adjustments are required.The dimensionality reduction strategy of.In this thesis,with the help of two di?erent computer vision tasks,the dimensionality reduction capability of unsupervised sparse self-coding is fully explored,and then the dimensionality reduction technology of unsupervised sparse self-coding is applied to enhance the application of object detection.The main research contents include the following aspects:1)Sparse autoencoder dimensionality reduction and its characteristics for data classification.In this study,the relationship between the number of iterations of unsupervised sparse autoencoder and the degree of output loss was explored and the corresponding recovered data was shown.In addition,unsupervised sparse autoencoder and other classic dimensionality reduction methods are used to reduce the dimensionality of several open data sets.Finally,the low-dimensional data after dimensionality reduction using unsupervised sparse autoencoder is classified to fully verify the good performance of unsupervised sparse autoencoder in data classification tasks.2)Sparse autoencoder dimensionality reduction for social image understanding.Images on social media are often tagged with a variety of non-professional labels.In response to this problem,this study uses unsupervised sparse autoencoder to reduce the dimensionality of social media image data to assist in the automatic annotation of social media images.At the same time,considering that multiple tags of social media images are relevant,this feature also provides a reference for the automatic annotation of social media images.Finally,it is verified that unsupervised sparse autoencoder is of application value in social media image understanding tasks.3)Sparse autoencoder dimension reduction for UAV object detection enhancement.Aiming at the problem of low target confidence and poor detection e?ect in the UAV object detection task,the HOG feature of the detected target is extracted and the low-dimensional HOG feature is extracted using unsupervised sparse autoencoder.Then use the classifier to perform secondary classification,weight the target confidence that the detection result is consistent with the classification result.Finally,a dual detection system suitable for UAV targets is proposed,the influence of confidence weighting on detection results is further discussed.