[Uridoki] AI Supports Product Category Selection, Simplifying Appraisal Requests

Uridoki, Inc. (HQ: Shinjuku, Tokyo; Representative: Yasuo Kogure), which operates the C2B resale platform "Uridoki" (https://uridoki.net/), announced the implementation of the "Uridoki Item Image Classifier," an AI model specialized in reuse products.

When a user uploads a product image for an appraisal request, this model analyzes the image features and calculates the probability distribution for each category. If the value for a specific category is significantly higher than all others, it classifies the item as that category. If the deviation is small, it guides the user by suggesting candidates for appropriate category selection.

This model is optimized for high-precision classification of images frequently found in the reuse industry, such as bottom or interior shots of brand bags and low-angle shots of high-end watches including bands.

The model was developed based on over 1.2 million real-world reuse product images accumulated by Uridoki, and supports 42 categories at launch. The overall accuracy for major categories in validation data is 87%. Furthermore, when limited to front-facing or full-view images registered by users, the classification accuracy exceeds 95%, confirming high performance in practical operations.

Development Background: Reducing Hesitation and Burden in Category Selection Uridoki is improving the appraisal request process using AI technology to realize a service experience where users can easily request appraisals. Previously, users had to manually select product categories. To reduce this burden, the company developed and introduced an image recognition AI model that automatically classifies categories from product images, aiming to reduce input workload and improve convenience.

Features and Advantages: Proprietary Training Specialized in the Reuse Domain On actual resale platforms, many images are posted showing only parts of products or items with signs of use and wear. While based on existing image recognition models, this model is uniquely optimized (fine-tuned) using Uridoki's proprietary transaction image data, enabling classification even for angles and conditions that are difficult for general-purpose image recognition models to discern.

System Processing Flow and Main Features The model operates through the following flow: - Product Image Upload: Users upload images of items they wish to have appraised from smartphones or PCs. - Automatic Image Analysis and Fast Processing: The AI instantly classifies based on the uploaded image data. - Optimal Category Assignment Based on Probability Distribution: Rather than forcing a single category, the AI calculates the probability distribution for each and presents the most likely one. - Proper Exception Handling for Difficult Images: If classification is uncertain due to distributed probabilities, the system quickly falls back to alternative processes, such as presenting a candidate list or asking the user to select directly.

Future System Development Uridoki continues to promote service improvements utilizing AI technology, having already developed and introduced AI systems for automatic detection of abnormal appraisal prices (1st wave), appraisal image analysis (2nd wave), and review moderation (3rd wave).