Construction And Application Of Classifiers With Limited Training Samples

Data-driven classifier construction methods highly rely on the quantity and quality of training samples.In recent years,along with the development of sensor,Internet and computer techniques,big data has been accumulated resulted from industrial production and our every-day life.Huge amount of data can be used to train complex classifiers with more parameters,thus improving the progresses in artificial intelligence and machine learning.However,for real applications in a specific domain like remote sensing image interpretation and mine safety analysis,or more complex pattern recognition problems including fine-grained classification,multi-label classification and semantic segmentation,it is difficult and expensive to collect,cleanse and label the original,massive samples.This is a restriction for the construction and application of classifiers in different real scenarios.For natural resource management,3 different classifiers are constructed and applied in fine-grained remote sensing image classification,multi-label remote sensing image classification and mine data classification respectively.It is valuable to design classifiers like these against the background that Ministry of Natural Resources is established and it will plan and manage land,minerals,forestry and other natural resources in China integrally.In this paper,“limited training samples" denotes the samples in small size or with poor quality in real applications.The classifier for scenarios like this can be implemented by improving deep networks by transfer learning,adversarial learning or other strategies on one hand.It is also feasible to process the data by expert knowledge,than construct classifier based on statistical learning theory for limited samples.The main work in this paper can be listed as follows.Firstly,a novel classifier based on transfer learning is proposed for improving the annotation and feature extraction issues in fine-grained categorization problem.We call the classifier attentioanl multi-adversarial networks(AMAN)and then apply it to handle fine-grained classification tasks in remote sensing interpretation.AMAN is able to utilize the coarse-grained labels in source domain for feature extraction initially.After that,with only a few samples in target domain,the classifier can enhance and align regional features from coarse to fine by attention proposal in adversarial learning.Experimental results and visualized analyses show that AMAN can deal with fine-grained classification successfully with incomplete,limited training set.Only need about 20%of the training samples from the complete training set,AMAN can obtain better accuracies than the mainstream algorithms in the past 5 years.Secondly,in order to extract object-level visual features and take a good use of label dependencies in multi-label tasks,a new classifier with CNN-LSTM structure is proposed,namely CM-GM(Cross-Modal Representation Learning and Label Graph Mining based Residual Multi-Attentional CNN-LSTM).CM-GM is used to deal with multi-label image recognition and remote sensing interpretation.At first,CM-GM generates label embeddings by text representation learning and label-graph mining which can excavate semantics in labels and utilize label dependencies effectively.After that,a channel-wise attention mechanism is introduced in CNN(Convolutional Neural Network)to extract object-level features.Based on the 2 steps above,we design a cross-modal alignment module to map the object-level image features and label text features in a shared space.These aligned visual embeddings can be used to train the LSTM(Long Short-term Memory)predictor.Due to the fact that CM-GM takes full advantage of text information and label dependencies,the classifier can make a good prediction with partial obvious objects in the picture.Our experiments show that CM-GM can obtain good accuracies,similar to state-of-the-arts,with roughly 60%labels in original data sets.The advantage is more obvious in data sets with complex label space and more labels.Finally,an improved LSSVM(Least-squares support-vector machines)based on adaptive artificial bee colony algorithm is proposed in this paper for mine safety rating and classification.We utilize LSSVM to classify the limited number of samples processed by experts.In order to optimize the hyper-parameters of LSSVM,an improved artificial bee colony algorithm(ABC)is presented.In the enhanced ABC,location updating formulas are adjusted adaptively according to the optimization process.Experimental results show that the proposed classifier has a good performance in accuracies with limited samples and small swarm population.On the other hand,in order to collect data about mine slopes,a new strategy which model the aerial photography and inspection mission of UAV(unmanned aerial vehicle)into a 3-dimensional travelling salesman problem which can be solved by ant colony optimization.Compared with manual control method,the strategy can reduce flight distance and collect more photos in limited time.