Shoe printing image identifies is an important application in computer vision in the work of public security. The current problems that cannot be accurately recognized in the public security investigation work restrict the improvement of work efficiency and quality due to the complexity of the shoe printing field extraction, the complex characteristics of the shoe print pattern and the incompleteness of the shoe printing images. In order to further improve the results of the printing of disabled shoes,a double-tower network shoe print retrieval algorithm based on feature screening was designed. On the one hand, a partitioning strategy was introduced into the network to divide the shoe printing image into the sole region and the heel region, and two feature networks were used to extract the image features respectively for fusion. On the other hand, convNeXt network, a new convolution neural network that integrates ResNet network and Transformer network, was selected as the backbone network to add attention mechanism module. After extracting the last layer of convolution features, different feature screening methods were used to remove the irrelevant features in the shoeprint image. Finally, the feature descriptor was spliced and expanded to calculate the similarity. In the training phase, the learning strategy was optimized, and it is trained as a complete image classification network. The experimental results show that the network model selected in this paper is better than other convolutional neural networks, and the CSS-200 and FID-300 two shoe print data sets have achieved high accuracy.