Research On Noisy Label Detection And Repair Network Model For Hyperspectral Image Classification

Hyperspectral images have rich target space and spectral information and have important applications in military and civilian fields.In recent years,great progress has been made in hyperspectral image classification tasks based on deep learning,however,due to the influence of noise labeling,the classification tasks face great challenges.In this paper,the problem of noise label detection and repair in hyperspectral images is studied from three aspects.Firstly,training the neural network in the noisy label dataset will overfit the data and seriously reduce the generalization performance of the neural network,so a noise label multi-level self-healing network model is designed.According to the spatial spectrum attention mechanism,a multi-level stacked empty spectrum attention feature extraction network is designed.The self-attention mechanism and self-healing module are used to design a detection and repair algorithm to detect and repair noise samples.In addition,when the neural network detects and repairs the hyperspectral data containing noisy labels,the detection and repair effect of small sample categories is not good,so the design and repair module re-repairs the data of small samples,which effectively reduces the noise rate and significantly improves the classification accuracy.Secondly,when the existing noise label detection algorithms detect samples,most of them use a single model for detection,which is less accurate than the binding force of multiple models on noise labels,and it is easy to misjudge when detecting noise label samples.Therefore,a plug-and-play self-integrated noise label detection and repair model is designed,according to the model output at different training stages,the average value of the prediction results of the overall training set is obtained,and the reliable and unreliable data sets are selected,and the next step is to use the self-healing module to effectively reduce the noise rate of the unreliable training set and significantly improve the classification performance of the neural network.Finally,the remote sensing image supervised classification technology requires a large amount of real label data to train the neural network,but marking samples of real label on hyperspectral images is time-consuming,expensive and requires field investigation,and the small number of labeled samples is an inevitable problem.Therefore,a semi-supervised noise label detection and repair model is designed to introduce unlabeled data.The self-integration model and self-training algorithm arecombined to realize the reliable selection and annotationof unlabeled data.ln order to reduce the impact of noiseon the model,the self-healing method of noise labels is added,which not only reduces the noise rate of the noise label training set,but also ensures the pseudo-label quality of unlabel data.lt also alleviates the problem of false label accumulation in self-training algorithms.Compared with the existing hyperspectral image classification model with noise labels,the effectiveness of this method is confirmed.