Research On The Rotational Invariant Deep Hashing Algorithm For Cross-Modal Retrieval

In recent years,everyone can use the network to publish information,transmit information and receive information conveniently at right time and right place because of the rapid development of mobile internet.This information typically contains a variety of data,such as text,voice,image,video,and so on.It is these rapidly growing data that make the application of retrieval to be increasing.Because of the large amount of data,the researchers have to consider the retrieval speed and the required storage space.The traditional method is inefficient on the large-scale dataset.And,the method must consume huge storage space which increase the cost of retrieval.With the development of research,the deep hashing model was proposed by scholars which can make the retrieval speed greatly accelerated and the storage space required for data was greatly reduced.Therefore,our work based on the current hot area of research.We fully studied the deep hashing algorithm and applied this algorithm to cross modal retrieval.In the process of studying deep hashing algorithm,we found that the rotation of images has influence on the feature extraction of deep hashing model.Using the rotated image as the input of deep hashing model which the extracted binary hash code have been changed.The change can reduce the retrieval accuracy.Therefore,this paper is devoted to the study of the deep hashing model with rotation invariance to change the present situation.The main work of this paper is as follows.We further studied the deep hashing algorithm.We found that the invariant properties of images have certain advantages for the feature extraction of image.The so-called rotation invariance is when the image is rotated to a certain angle,it still does not affect the retrieval results.Therefore,we added a rotational invariance loss function and applied it to the image modal feature extraction in the deep neural network to restrain the difference of rotated image features and original image features.Thus we minimize the influence of the rotation.In the image preprocessing stage,we rotated the images according to-10,-5,0,5,and 10 degrees.The advantage of this method is to make full use of the image information.It can not only make the image feature to be invariant to the rotation and local noise,but also enhance the robustness of the network.Finally,we applied the rotation invariant deep hashing model to cross-modal retrieval.And then,we make use of the model to extract features of image and text.We get better results.According to the experimental results,we can see that the accuracy of the retrieval is improved.