An efficient network based on double constrained loss for fabric image retrieval

Abstract

In order to efficiently retrieve the same or similar fabric samples, a fabric image retrieval model based on deep hashing is proposed. The model is improved on the basis of MobileNetV1, and the performance of the model is improved by combining the h-swish activation function and attention mechanism module. The hashing method is used to solve the problem of low feature matching speed caused by high dimensional feature output. By combining the label information and similarity information of the images, a new loss function is constructed, which solves the problem of low model accuracy caused by the large feature difference between similar samples and the small feature difference between heterogeneous samples in fabric images. The experimental results on self-built fabric image dataset showed that the feature extraction time of the proposed algorithm was 0.25ms, and the MAP reached 93.2%, which can take into account the fabric retrieval speed and improve retrieval accuracy at the same time, and has certain application prospects.