Image Recognition Based on Compressive Imaging and Optimal Feature Selection

The measurement matrix in compressive imaging controls the crucial feature information for high performance recognition. In this study, a deterministic orthogonal measurement matrix design method using the discrete cosine transform and a compressive feature selection scheme are proposed to implement high-end computational optics for imaging. The selection scheme systematically evaluates the recognition importance for the frequency features, combined with a scaling of the contribution of the various coefficients used to produce a base matrix for the new group of measuring patterns, which ensures the minimal recognition difference for each individual order of frequency filters and combining a relatively complex expression to quickly find the best quantization values. The model parameters are gradually adjusted and eventually converge to the best result through training with a large number of pre-determined samples from the dataset and backpropagating the feature selection loss along with the recognition loss, and the data processing capabilities can be enhanced because the measurement matrix is a priori information for the recognition phase. The systematic ability of the proposed technique was verified through simulations and experiments on two standard datasets: MNIST and CIFAR-10. The results show that the proposed method outperforms state-of-the-art methods in terms of both the model complexity and classification accuracy, which indicates that our study provides a new practical solution for compressive imaging recognition.