Luxury Goods Identification based on Multi-Head Convolutional Neural Network with Attractor Loss

The counterfeit goods trade is a global issue. Traditionally, customized equipment, elaborate labeling, and expert inspection are the primary ways to identify fakes. In recent years, deep learning technology has been used for counterfeit identification based on image detection and classification. There are two critical challenges. First, in real-world situations, the counterfeiting patterns are diverse, while the genuine ones are often identical. This is because counterfeit goods may be partially assembled from some parts of genuine products. Second, authenticating a counterfeit as genuine can cause a lot of trouble. To address these issues, we propose a novel end-to-end algorithm that combines a new Multi-Head Convolution Neural Network (MH-CNN) with a new Loss function named Attractor Loss (AtL) and applies it to image classification of real and fake goods. Technically, our method takes multiple images of different positions of an object as input. Then it outputs 1) the classification probability of belonging to a real class, and 2) a feature vector that is used for nearest neighbor retrieval. On engineering implementation, we integrate an MH-CNN structure, which contains a shared CNN backbone for extracting deep features and multiple heads for computing the deep representation of different positions. To improve the recognition performance, we propose a joint training method to train the MH-CNN model and propose a new AtL to filter vector representation of images from real class and pull them closer for enhancing the recognition rate without loss of accuracy. Extensive experimental results demonstrate the superior performance of our proposed method.