Blur Retail Product Images for Fine-grained Image Classification Combining Text Area Features

There are small differences between different retail product images belong to the same category, such as product models, sizes, and tastes. Retail product images do not have target local regions which can be used for detection and classification. Besides, the differences between subcategories are more subtle. These make fine-grained image classification for retail product images more challenging. It is hard to locate discriminative local areas in the product images, which are widely used in training for specific area detection such as head and trunk in fine-grained classification of birds images. However, the existing fine-grained image classification for product images only use image features to target local regions, but ignore text features in the images. Product images contain rich text information. Text feature is a very important component in product images. Since text recognition methods do not achieve satisfactory results on retail product images of low quality, we consider introducing text area features into fuzzy product image classification. We propose a fine-grained classification method for blur retail product images combining the features of text regions. First, we use the density-based clustering method to find category key areas in the text regions detected in the product images. Secondly, combine the key text areas features in the image when doing fine-grained classification. Experimental results show that our method achieves a higher classification accuracy in blur retail product image classifications, getting 4% improvement in the original classification results.