Discriminative Region Guided Deep Neural Network Towards Food Image Classification

Food image classification plays an important role in smart health management, such as, diet analysis and food recommendation. Due to the similar appearance and shape between different foods, it is quite challenging to distinguish various food categories from their images. To address this issue, we propose a discriminative region guided deep neural network to classify the food images. More specifically, a saliency map based pooling strategy is applied to the input image to preserve the category aware discriminative regions. Meanwhile, the multi-scale fusion scheme is employed in our deep neural network to describe the discriminative regions across different resolutions. Experimental results on a large-scale Chinese food database show that, the average accuracy the proposed method is as high as 91.18%, and outperforms the baseline by 2.58%.