Semantic-Based Interactive Image Retrieval

With the advancing of research on image retrieval technology,its application scope is more and more extensive.Combining image retrieval with deep learning and interactive technology has gradually become a research hotspot.At present,there are still many areas for improvement in image retrieval technology for specific data sets.For example,improving the segmentation performance of an image,feature extraction for specific data,and effective use of user feedback.This thesis mainly focused on the research of skin disease image retrieval,and introduced transfer learning and related feedback to obtain better retrieval results.In this thesis,image segmentation,classification and retrieval of pigmented skin disease data are carried out.The main research contents and results are as follows:1.Introducing the user’s will,using DeepMask interactive segmentation algorithm to effectively segment the skin disease image.DeepMask is used to train the segmented mask of skin disease image,and the user clicks to update the weight of the mask to obtain the desired target area.2.Using transfer learning to classify images of skin disease small datasets from mobile phones/cameras.The InceptionV3 model is pre-trained through a large-scale similar data set to obtain initial parameters,and then the model is fine-tuned to complete the multi-classification of the skin disease image,and the semantic label of the image is obtained.A comparative study of different fine-tuning methods to obtain a solution suitable for the skin disease data set.3.An interactive retrieval algorithm based on deep learning and SVM related feedback is proposed.Because the traditional feature extraction method cannot describe the characteristics of skin disease images well,this thesis uses convolutional neural network to extract image features,and through SVM-based correlation feedback algorithm to continuously learn the user’s search intention and optimize the search results.4.An evaluation model of interactive retrieval algorithm based on the number of user feedback is built.By measuring the number of user interactions to evaluate the performance of the interactive retrieval algorithm,it is possible to better measure the sensitivity of the algorithm to user feedback.