Deep Convolutional Neural Networks for Near-Duplicate Image Detection

Near-duplicate image detection has become important and widely used in the era of big data. This category of task can be very challenging for deep learning methods due to a number of reasons. Firstly, because most of the data available is noisy and can be hard to clean, costing a vast amount of time and effort to generate a big dataset to feed the deep neural network. Moreover, it is crucial to select a well-suited architecture and optimized training that handles the input data well and outputs rightful solution. In this project, I developed an end to end deep convolutional neural network (CNN) to address the problem of recognizing the near-duplicate images in a set by producing a spatial representation for each image. The work is implemented based on a similar procedure in the state-of-the-art literature by Gordo [1]. The network is coded in CAFFE framework using a Siamese architecture of ResNet-101. The results are promising, demonstrated by a constant decrease in the loss value of the ranking function of the network when fine tuned by borrowing initial weights from the ResNet-101 model trained on ImageNet by Kaiming He [2]. A. Gordo, J. Almazán, J. Revaud, and D. Larlus, “End-to-End Learning of Deep Visual Representations for Image Retrieval,” Int. J. Comput. Vis., vol. 124, no. 2, Sep. 2017. Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun, "Deep Residual Learning for Image Recognition", arXiv preprint, arXiv:1512.03385, 2015.