Article Preview
Top1. Introduction
With recent developments in deep learning technique, an increasing number of studies are being conducted on images and image recognition (Al-Halah, Z., Stiefelhagen R., Grauman K. 2017; Mihm, B. 2010; Lee, H., Seol J., Sang-goo Lee. 2017). Among these, efforts are focused on methods of searching clothing product images, that can identify fashion trends and clothing features. Such a system should be capable of analyzing the fashion trends of many clothing products quickly. In the present study, a clothing product image search system is proposed and its performance evaluated. The system collects clothing product images from about 300 clothing-related shopping malls on a real-time basis, rapidly classifies the clothing images relative to a new input image, and analyzes the trend in clothing products.
Figure 1.
Similar image results of clothing
Figure 1 shows an example of a similar image search for the ‘S’ Shopping Mall. The images in Figure 1 were searched for specific clothing product patterns, poses and tags. A feature-based analytical technique is employed. The deep learning model provides a simple comparison of characteristics, stable service while handling numerous products, rapid updating of new products, and can process a large number of search inquiries.
Establishing a deep learning system for precise searches based on a huge number of images can be high cost, which small companies can hardly afford. Many recent studies on image searches have been conducted using Convolutional Neural Network (CNN). However, similar image analysis with CNN is performed by exploring all the pixels of an image, and thus requires countlessly repeated learning procedures, and a long time to measure similarity (Lee, S., Choi M.-H, Lee H.-J., et al., 2019; Seo, M., Lee K.-Y., 2019). In addition, if the frequency of the repeated learning procedures is decreased to increase computation speed, the level of similarity is reduced.
To overcome these problems related to CNN, this article proposes a model for similar image searches based on the Artificial Neural Network Oh Yeah (ANNOY) using a binary tree, with a little trade-off in image similarity (Bernhardsson E., 2020). The image similarity and the image search measurement time using the CNN and the ANNOY were compared. For the performance comparison, a clothing product dataset was constructed to collect images from about 300 clothing-related shopping malls. The similarity among the fashion products were compare and analyzed using the two techniques. Python was used to collect and update the images of fashion products from about 300 shopping malls on a real-time basis. The image dataset was constructed using 60,000 collected images: 40,000 images (67%) were used for transfer learning, and the other 20,000 images (33%) were used as the input images for the similar image search. The images’ similarity and measurement times for the two techniques were analyzed and compared.