How Metadata Automation Helps Physical World Retailers Compete with Ascendant E-Tailers
2018-05-22
By increasing the transparency of product information, established retailers can help consumers make more informed purchases and better compete with more digitally adept e-tailers.
With online sales growing faster than ever, traditional retailers are rightfully investing in omnichannel strategies and redoubling their efforts to meet digitally inclined consumer demands. Offering more detailed product information, images and sometimes even video is one way the establishment can help assure and empower buyers and regain their competitive edge versus pure-play e-tailers.
To enhance online product discovery, retailers must maintain and provide digital images and videos, catalog descriptions, category-specific metadata (such as nutritional information for food products), stock availability, product size ranges, product ratings and reviews, pricing, and promotional information for all physical stock keeping units (SKU). Acquiring this information from suppliers is usually a time-consuming task, requiring various handoffs and significant manual labor.
To address this challenge, we built an intelligent automated system that extracts more searchable retail information from actual product labels. Using computer vision, natural language processing, and machine learning techniques, the system can extract important metadata such as product title, product description, volume/weight, nutritional facts, branded logos, and barcodes.
Test results in our labs show 95% accuracy, which is good news for retailers hoping to get their house in order to better compete with metadata rich online stores. Here’s a look at how our technology works and why it’s important.
The Metadata Disadvantage
When it comes to supplying complete and accurate product information, many retailers are at a severe disadvantage. This is because stores typically rely on suppliers to provide product images and metadata through various methods (electronic data interchange, printed or digital catalogs) and various formats (text, Excel, PDF, XML). Sometimes retailers purchase product information from third-party providers to fill in the gaps. Either way, this often results in an incomplete or otherwise underwhelming experience for prospective buyers who, in search of better information, usually take their business elsewhere.
Established retailers also face several technology challenges when trying to extract content from retail product labels. These include region segmentation, diverse product backgrounds, typography and font usage, cursive or handwritten text, lighting conditions, camera artifacts, and low-quality images. Other obstacles include variations in product size, incongruent metadata placements (either on box or product itself), and poor information alignment.
And yet, next to price, high-quality product images and rich metadata are a primary driver for consumer purchases. According to Brand View, 98% of shoppers expect retailers to show comprehensive product data. In fact, there is an indisputable link between the quality and completeness of online product content and sales. To compete, retailers must consider a better alternative.
Figure 1
Automation to The Rescue
This much we know: when retailers improve the accuracy and completeness of online product content, sales increase and returns decrease by as much as 4%, according to a recent MarketWatch study.
Our five-step solution can ease the extraction of product metadata from existing retail labels. In short, the solution photographs a retail product carton from all sides. The captured images are then fed into an algorithm that performs data extraction. Image pre-processing techniques are then applied to identify and catalog important information, including include brand name, product name, logo, food certification logos, net weight, nutritional facts and barcodes.
This is done with computer vision technology that removes background colors and provides image quality checks. It applies optical character recognition (OCR) and algorithms that provide text region detection, as well as auto-detection technologies that surface net weight and nutritional facts.
To assess the performance of our proposed solution, we evaluated a data set of 352 food products from 53 different brands and 955 product images to achieve 95% accuracy.
The Road Ahead
Product labels are a trusted source of metadata information. Hence, our automated extraction solution reduces the burden of validating product data from various vendors and provides the additional information that today’s digital customers demand.
All told, computer vision and automated extraction show clear potential for improved product metadata search and supply. We believe our proposed computer vision approach can be extended to other product categories, such as health, beauty, books, toys and video games.