Deep Image Feature Learning With Fuzzy Rules

Image feature learning is a basic research topic in the field of computer vision and machine learning.The various tasks in computer vision,such as image classification,object detection,and scene segmentation,all treat feature learning as the initial step using various feature learning methods,which is followed by other techniques to achieve their goals.According to existing research,the popular image feature learning methods are learning subspace and deep neural networks,which can be automatically extract robust features through end-to-end training instead of hand-crafted feature extraction for classification tasks.However,the subspace learning methods have shortcomings in the choice of the kernel function.Deep neural networks also currently face some challenges.Their effectiveness is heavily dependent on large datasets,and they are usually regarded as black box models with poor interpretability.To meet the above challenges,this paper conducts research in the field of image feature learning.The main works of this paper are as follows:1)The first work is based on the TSK fuzzy system,a model of image feature learning with fuzzy rules is proposed,which has the strong learning ability and good interpretability of the fuzzy system.Different from classical fuzzy system,which is usually used for classification and regression tasks,TSK fuzzy system is regarded as a feature extraction model for image feature extraction.The antecedent part of the TSK fuzzy system is regarded as a non-linear transformation that maps the image data from the original space to the high-dimensional space,and the consequent part of the model can choose different feature learning target criteria according to the needs,which avoid the choice of kernel function and also makes the model have the ability to extract features non-linearly.2)The second work is to explore the depth of the proposed model structure,an image depth feature learning method based on fuzzy rules is proposed.This method progressively extracts the image features by stacking multiple layers of TSK image feature learning modules,which can effectively combine the advantages of fuzzy systems with interpretability and the features of progressive extraction of deep learning architectures.More importantly,the learning process of the method is only based on forward propagation without back propagation and iterative learning,which results in high learning efficiency.In order to maintain the structure of the image data during feature extraction,the principal component analysis(PCA)is used to maintain the geometric properties of the data in the feature learning stage of each layer.3)The Third work is to explore the width of the model structure,a multi-grained image feature learning method with fuzzy rules is proposed.The multi-granularity window scanning strategy is adopted in this method to integrate the features extracted from the TSK-FS image feature learning in multiple granularities,which further improves the feature learning ability.In order to retain the discriminative information of the original data as much as possible,so that the data has the best separability in the new feature space,linear discriminant analysis(LDA)is used to maintain the discriminative information of the data in the feature learning stage.Extensive experiments are conducted on image datasets of different scales with two proposed methods.The experimental results clearly show the effectiveness of both methods,and have strong robustness to interference factors such as noise,occlusion and lighting.