Image Retrieval Based On Deep Learning

With information of Image processing becomes more and more massive, Text-based Image Retrieval(TBIR) has been increasingly anachronistic, so researchers gradually turned their attention to the technology of Content-based Image Retrieval(CBIR) and the rise of various algorithms about Content-based Image Retrieval becomes rapidly. It is a significant task for us to describe content of images effectively on large-scale content-based image retrieval. Because the binary hash code has a high-efficiency in calculation and storage, the binary hash algorithm has caused the extensive concern, but at the same time extraction SIFT description of midnight, will reduce the computational speed of the algorithm. Among these technologies, as one of the most popular technology in machine learning research, deep learning can let model directly study characteristics of the images through building the mode, and thus it is greatly reduce the deviation of artificial to extract the image feature. As the neural network in deep learning is relatively simple and efficient, and has more accurate in characteristics extraction than traditional extraction algorithm, it is became one of the most popular technology currently. Convolution neural network(CNN) has made a major breakthrough in the field of image retrieval. The use of CNN can make the model not only learn binary representation of the training sample data the through the hidden layer, but also learn image representation.First of all, the paper puts forward an effective deep learning network model. With this model, we can generate a binary hash code and use it for quickly image retrieval. In addition, the network model can not only describe relevant images of the fields, but also to learn a series of hash function through adding a hidden attribute in neural network model. As we known, most of the supervised learning algorithms usually require two images as input to learn the binary representation of the image, but the deep learning methods of the paper only need one image as input to study image binary encoding and image representation. In this way, it greatly reduces the computation work and storage space. Therefore, the depth study of binary hash algorithm of the paper is especially suitable for large-scale image retrieval.Secondly, the retrieval rate is very high when using binary features for image retrieval, On the other hand, the retrieval precision is high when using floating point features for image retrieval, but it needs large amount of calculation work and the retrieve rate is very low. In order to solve this problem, we put forward a comprehensive from coarse to fine retrieval method. We will use binary feature to search out the similar picture of query image. Because a binary code can correspond to a lot of pictures of different labels, there are a lot of interference in pictures among retrieved images by using of binary features, this step is a coarse level of retrieval. And when we use floating-point features for further eliminate interference image, this step is fine level of retrieval. In this way, we can not only ensure the precision of image retrieval, but also improves the speed of the image retrieval.Through on the data sets of different sizes in three kinds of experiments, model test based on the MNIST data set, model test based on the CIFAR-10 data set, model test based on the Yahoo-1Mdata set. We can draw the following conclusions, with the increasing of data sets, the model of the paper still maintained a highly efficient and stable performance. This shows that our model is more suitable for large-scale image retrieval.