Image feature extraction based on fuzzy restricted Boltzmann machine

•Sparse penalty factor is used to control the hidden element of the model.•The hidden layer constraint based on fuzzy number is introduced into RBM model.•Without affecting feature learning, a de redundancy mechanism is introduced.•Research on adaptive step size algorithm based on error representation. In order to accelerate the speed of visual image feature extraction and solve the problem of training deceleration caused by the introduction of fuzzy numbers into the original Restricted Boltzmann Machine (RBM), a Mixed Accelerated learning method based on a Fuzzy Restricted Boltzmann Machine (MAFRBM) is studied. Firstly, the penalty factor is introduced to increase the sparsity of data and improve the effectiveness of feature information extraction. Secondly, the characteristic similarity of the hidden layer units of the network is compared. When the unit similarity is higher than the threshold, the unit is a redundant unit for repeated learning, which is eliminated to accelerate learning. Thirdly, the reconstruction error of each training is used to adaptively change the network learning rate to achieve further acceleration. Finally, compared with the change of network reconstruction error, the accuracy of network convergence is defined to accelerate learning convergence. And a multi-functional image recognition model is constructed, which is MAFRBM + SVM. The features extracted by MAFRBM unsupervised learning are imported into SVM to test the effectiveness and practicability of MAFRBM. The experimental results show that the learning speed and feature extraction ability of MAFRBM are better than other models.