Image-free target recognition approach based on single-pixel imaging is a novel strategy for target recognition. It normally utilizes convolutional neural networks to directly recognize the detected sequences acquired by single-pixel detectors. This process does not reconstruct the image, which greatly reduces the recognition time and storage consumption. However, the sub-Nyquist single-pixel detections loss some spatial information, resulting in low recognition accuracy. To solve this problem, we propose a new image-free recognition method called encoded wide residual network, which combines the encoded matrix and the improved wide residual network. The method introduces a small-size modulation matrix to obtain a sequence of single-pixel probe value matrices and applies a Kronecker product operation with the modulation matrix to obtain a new sequence of encoding matrices. Using the sequence of encoding matrices as an input to the neural network effectively drives neural network learning. This method reduces the number of model parameters and further improves the accuracy of image-free recognition. Simulation and experiments were carried out on the Fashion MNIST dataset, and the results show that the method improves the accuracy at different sampling rates such as 3%, 5%, 9.77% and 25%.