Deep net architectures for visual-based clothing image recognition on large database

In the Big Data era, there is a need for powerful visual-based analytics tools when pictures have replaced texts and become main contents on the Internet. Hence, in this study, we explore convolutional neural networks with a goal of resolving clothing style classification and retrieval tasks. To reduce training complexity, low-level and mid-level features were learned in the deep models on large-scale datasets and then transfer learning is incorporated by fine-tuning pre-trained models using the clothing dataset. However, a large amount of collected data needs huge computations for tuning parameters. Therefore, one architecture inspired from Adaboost is designed to use multiple deep nets that are trained with a sub-dataset. Thus, the training time can be accelerated if each net is computed in one client node in a distributed computing environment. Moreover, to increase system flexibility, two architectures with multiple deep nets with two outputs are proposed for binary-class classification. Therefore, when new classes are added, no additional computation is needed for all training data. In order to integrate output responses from multiple nets, classification rules are proposed as well. Experiments are performed to compare existing systems with hand-crafted features. According to the results, the proposed system can provide significant improvements on three public clothing datasets for style classifications, particularly on the large dataset with 80,000 images where an improvement of 18% in accuracy was recognized.