Image retrieval and classification on deep convolutional SparkNet

Image retrieval and classification are hot topics in computer vision and have attracted great attention nowadays with the emergence of large-scale data. We propose a new scheme to use both deep learning models and large-scale computing platform and jointly learn powerful feature representations in image classification and retrieval. We achieve a superior performance on the ImageNet dataset, where the framework is easy to be embedded for daily user experience. First we conduct the classification task using deep convolutional neural networks with several novel techniques, including batch normalization and multi-crop testing to obtain a better performance. Then we transfer the network's knowledge to image retrieval task by comparing the feature codebook of the query image with those feature database extracted from the deep model. Such a search pipeline is implemented in a MapReduce framework on the Spark platform, which is suitable for large-scale and real-time data processing. At last, the system outputs to users some textual information of the predicted object searching from Internet as well as similar images from the retrieval stage, making our work a real application.