Improving deep convolutional neural networks for real-world clothing image

Clothing images are abundant especially from the ecommercial platform, due to the rapid development of e-business. Recognizing and retrieving those images is of importance for commercial and social applications, which has recently been received tremendous attention from multimedia processing and computer vision. However, the large variations in clothing of their appearance and style, and even the large quantity of multiple categories and attributes make those problems challenging. Furthermore, the labels provided by shop retailers for real world images are largely erroneous or incomplete. Even worse, the imbalance problem among those image categories prevents the effective learning. To overcome those problems, we adopt a multi-task deep learning framework to learn effective representation. And we propose multi-weight convolutional neural networks for imbalance learning. The topology of this network is composed of two kinds of layers, shared layers at the bottom and task dependent ones at the top. Furthermore, category-relevant parameters are incorporated to regularize the learning procedure of backward gradients for different categories. We collect a large-scale dataset for those two problems containing about one million shop photos from four different Chinese retailers. Experiments on this dataset demonstrate that our proposed joint framework and multi-weight neural networks can effectively learn robust representation and achieve better performance.