Shulou(Shulou.com)06/02 Report--

Since the retail field constructively put forward a new retail blueprint based on the three cores of "people, goods and market" around the transformation and upgrading in recent years, new technologies have flocked to compete in various types of retail scenarios. In the previous sharing, we mainly introduced how the singularity cloud cuts from the "human" dimension into the retail scene through artificial intelligence technology. This paper will focus on how the artificial intelligence technology of singularity cloud empowers retail business from the dimension of "goods".

As the first step in the landing of singularity cloud AI, visual intelligence engine plays a major role in commodity digitization. From a visual point of view, the morphological difference of "goods" has increased significantly compared with "people". The products developed by different vertical industries to meet the market demand are different in function and form, and even the same type of products will produce different forms to meet the differentiated demand. This not only poses a new challenge to the visual intelligence engine, but also stimulates a greater technological breakthrough for StartDT AI Lab. Below, several representative industry scenarios are selected from the scenarios where Singularity Cloud has been commercially landed to show the technological breakthroughs that StartDT AI Lab has made.

AI recognize goods and empower clothing industry

We took the lead in 17 years to launch clothing identification technology, combined with the product-Singularity Magic Mirror to achieve landing. In this scene, after the user stands in front of the singularity mirror for a few seconds, the singularity mirror will first divide and understand what the user is wearing, separating various types of clothing such as T-shirts, trench coats, jackets, jeans and so on. then through the analysis of individual clothing, such as T-shirt, we can extract features and analyze the T-shirt style, sleeve length, version and other attributes. Finally, in our self-built million-level clothing commodity library, we use the recommendation algorithm to recommend similar and matching goods for users, so as to achieve the effect of customer diversion and intelligent shopping guide.

On this link, the main problems and solutions we encountered at that time:

✨ 1. When obtaining users' accurate clothing information and features, obvious clothes such as T-shirts and shorts are easier to obtain, but longer clothes such as dresses and trench coats are more difficult to deal with.

By collecting a large amount of data, cleaning and marking, improving the data set, and using the cascade method, we first detect and analyze the clothing first-level label, and then carry out secondary processing, so as to improve the segmentation effect.

✨ 2. Clothing attributes are diverse and there is no unified standard, so how to judge quickly and at the same time becomes a difficult problem.

We make the classifiers of attributes independent of each other, and use features as inputs to reduce repeated operations. In addition, we connect the classifiers to Graph at the same time to make the whole process End-to-End, fast and effective.

✨ 3. Large-scale retrieval: when the database is large, the retrieval speed is slow and can not respond quickly.

We deploy the comparison database on the distributed cluster and implement the map-reduce at the feature comparison level, which makes us able to cope with various levels of comparison.

AI recognizes goods and enables quick consumption of beverage industry

We have been exploring the fast consumer drinks industry for nearly two years. A variety of drinks-including mineral water, carbonated drinks, fruit juices, beer and spirits-can be sold in the Singularity Cabinet. Through the self-developed deep learning algorithm, data sampling method and our customized hardware configuration, we have achieved more than 99% recognition accuracy in commercial scenarios. We control the damage rate to less than 1%, which is lower than the average level of goods loss in the fast consumer drinks industry. In addition, we have achieved a very high flat share in less than 1 square meter of space, and the monthly retail sales of a single cabinet can reach the level of 1000 yuan, but only a small amount of maintenance work is needed.

Behind the official commercial use, the main contribution of StartDT AI Lab has the following points:

A Fast and effective algorithm for small object Detection in No.1

Unlike most scenes, our detection targets are often small targets with the smallest edge close to 16 pixels, and there are quite a few targets in a single image. We use the adaptive anchor method to make the prior anchor more accurate; in addition, we use a feature enhancement algorithm to reduce the loss of detail features in the deep network as much as possible, so as to avoid the instability of low-score detection results caused by too few features. In addition, we have also developed a self-distillation method, which improves the accuracy of the model without increasing the parameters, thus reaching the commercial level.

Recognition Technology based on the combination of No.2 and Metric Learning

Measurement learning has been widely used in face recognition models in the past few years, and achieved very good results. We introduce it into commodity identification, combined with the classical neural network method, to make the recognition results more accurate and reliable; in addition, the model can also output comparable features, support the results of feature comparison, and support similarity comparison between different classes. thus at the level of product selection, it can avoid the problem that similar goods can not be identified when selling similar goods at the same time.

No.3 small dataset enhancements

Our dataset is actually a small dataset relative to the usage scenario. How to use small datasets to achieve commercial accuracy in big data scenarios? We have developed a mix method, which makes the detection model have a very high recall rate; in addition, we also use GAN to train the generator while training the classifier, and then use the generated results to train the classifier at the same time, so that the classifier is trained more fully and smoothly.

In the vertical field of fresh / medicine, we also dare to explore and try to give full play to the advantages of our own deep learning algorithms and sampling methods, combined with the advantages of self-developed hardware, to create intelligent containers that can make many kinds of fresh unaffected by categories. As we all know, in the vertical field, the appearance of the same sku varies, and the adaptation cost is high. Through specific product design, we perfectly support all kinds of fresh vegetables and fruits, and the accuracy can be close to 100%. So that customers are not limited to vegetable markets / pharmacies and other special venues, but also can give full play to the perfect advantages of unmanned containers-not limited by time.

(identification of drugs and fresh products from top to bottom)

The harvest and innovation of StartDT AI Lab's exploration in the dark:

Thinking about the novel Deep Learning scene of No.1

In order to get rid of the similarities and differences and incompatibility brought by various shapes and sizes under the fresh scene, as well as the new practical needs of customers quickly, we combine the characteristics of the goods, the advantages of the products and the principle of deep learning algorithm on the actual road of landing of the scene, to carry out special packaging for grotesque fresh products, this scheme solves the differences between fresh products, and can support customers to update quickly.

No.2 's unique and innovative data augmentation

Novel scheme design alone is not enough for deep learning to land perfectly in a real scene. Accuracy is the first element of customers, and data is the capital of deep learning. After deeply analyzing the data characteristics and algorithm performance of vertical fields such as fresh and medicine, StartDT AI Lab creatively broadens the data on the sampled data set, makes the accuracy almost perfect, and explains the concept of landing deep learning.

From the above cases, it presents the digital exploration of singularity cloud in "goods". Although it is still a relatively preliminary stage of commodity digitization, it also realizes the full-link digital tracking in the commodity sales cycle for the first time through AI technology. And on this basis, the retail front-end cost reduction is realized, and the efficiency of user research automation is improved. In order to make the business smarter, StartDT AI Lab will continue to move forward, please continue to follow ~

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.