Research On Fabric Image Retrieval Based On Convolutional Neural Network

With the vigorous development of electronic commerce,consumers' demand for clothing products has increased,which has made the market place higher requirements on textile fabrics,which is one of the raw materials for clothing products.In order to respond to the market,textile fabric manufacturers are committed to fabric innovation and research.The use of a variety of processing methods and materials for fabric production,which has also led to more and more types of textile fabrics on the market.For textile fabric manufacturers,the continuous growth of fabric types has brought pressure on inventory management.In order to facilitate the management of fabric inventory,this paper introduces convolutional neural network technology into fabric image retrieval,and makes the following research:(1)A fabric image feature extraction algorithm based on Inception V3 migration learning is proposed.Aiming at the problem of low accuracy of fabric image feature extraction due to various types of fabric images and complex visual features,the Inception V3 convolutional neural network transfer learning method is used to train the fabric image classification model,and the fabric image data set can be extracted through this model to describe the image features.The feature vector of the fabric is constructed to construct the fabric image feature database.In the fabric image retrieval experiment,Euclidean distance is used to calculate the similarity between the feature vector extracted from the query image and all feature vectors stored in the feature database.(2)A fabric image retrieval speed-up algorithm based on the improved Inception network is proposed.In view of the problem of slow retrieval speed caused by the large scale of fabric image sets,the Inception V3 network is improved by adding a hash coding layer and optimizing the loss function,and the fabric image classification model is retrained.The fabric image data set can be extracted through this model to describe the characteristics of the fabric image.The feature vector in the form of hash encoding is used to construct the fabric image feature database.In the fabric image retrieval experiment,the Hamming distance is used to calculate the similarity between the feature vector extracted from the query image and all feature vectors stored in the feature database.(3)Combining the two algorithms proposed above,a hierarchical retrieval algorithm is proposed,and a fabric image retrieval system for hierarchical retrieval is realized.First,use the two algorithms proposed in the previous article to jointly build a fabric image feature database,store two feature vectors of each image,for the query image,first use the fabric image retrieval speed-up algorithm based on the improved Inception network to perform a rough level of retrieval.The result of the retrieval is then used for the fine-level retrieval of the fabric image feature extraction algorithm based on Inception V3 migration learning,and the final retrieval result is obtained.Experiments have proved that the fabric image feature extraction algorithm based on Inception V3 migration learning has a higher retrieval accuracy than the commonly used fabric image feature extraction algorithm;the fabric image retrieval speed-up algorithm based on the improved Inception network has significantly improved the retrieval speed;hierarchical retrieval The fabric image retrieval system combines the advantages of the two algorithms,combines high retrieval accuracy and excellent retrieval speed,and has certain practical significance.