Design Of Printed Fabric Image Retrieval System Based On Deep Learning

In recent years,with the development of Internet technology,and the popularity of photographic hardware,images in all walks of life have shown explosive growth.It has become more and more difficult to find images that meets real needs from a massive data.As the largest textile producing and exporting country in the world,it's very important for China to provide a fast and efficient image retrieval system in the process of fabric design,inventory management and sales.In view of the outstanding achievements of deep learning in image recognition,detection and classification,a printed fabric image retrieval system based on deep learning is designed,which enables to help users to obtain the required information of fabric quickly.The main works of this thesis are as follows:(1)In view of the needs of textile industry for fabric image retrieval,the printed fabric image retrieval system is designed based on the functional and performance requirements,including hardware selection,platform construction and software module design.(2)Aiming at the problems of weak learning ability,slow speed and low precision in traditional text-based and content-based image retrieval,a method of printed fabric image retrieval based on convolutional neural networks(CNN)is proposed.In this thesis,the local response normalization(LRN)in AlexNet is substituted with the batch normalization(BN).Transfer learning is used to train the modified model,which not only can make up the lack of datasets in model training,but also can prevent overfitting effectively and promote the convergence rate of model.During retrieval,firstly,the trained model is used to extract the features of fabric.Then,similarities of images are measured by Euclidean distance.Finally,the top k images and the category,inventory code and popular year of the query image,etc.that user needs are returned based on the similarity.By comparing the experimental results,the proposed method obtained a better retrieval performance,and its average precision and average recall are reached 93.4% and 46.7%,respectively.(3)Aiming at the problems of time-consuming and memory usage of large-scale image retrieval,a method of printed fabric image retrieval based on hash coding and CNN is proposed.A coarse-to-fine search strategy is used.Firstly,the Hamming distances are calculated according to the binary hash codes of images to achieve the coarse search,and m images with high similarity are mapped into a new hash pool.Then the Euclidean distances between the fc7 features of the m images and query image are calculated for fine-level search.The experimental results show that the proposed method can combine the strong ability of self-learning of CNN and the advantages of hash algorithm in calculation and memory occupancy simultaneously,which is more suitable for retrieval of large data and has a certain practicability.There are 34 images,14 tables and 68 references in this thesis.