As an important part of people's livelihood consumption, fresh food retail plays a vital role in promoting the upgrading of new consumption in the industry. In unmanned retail scenarios such as supermarkets, the current settlement methods mainly include the following:
However, the above methods have the following disadvantages:
- Barcode method: It is relatively mature for finished packaged goods, but the fresh products themselves have "no code to scan";
- RFID method: the additional costs required are difficult to bear for a long time;
- Fresh weighing method: It is difficult to distinguish commodities of the same variety with different unit prices, and manual identification of commodity varieties is required, which is inefficient.
Therefore, it is particularly important to choose an identification system that can support various commodity identification on a large scale, facilitate management, and maintain low maintenance costs.
In response to the above industry needs, Paddle and Intel have provided a set of fresh and retail product recognition systems based on PaddleClas, an image classification development kit and OpenVINO™ tool suite, to meet the needs of product recognition scenarios that require scalability. In this solution, only one set of models needs to be trained. In the subsequent use process, there is no need to retrain the model frequently. It is only necessary to configure a small number of representative new product images in the retrieval library, which can well solve the problem of new products. Commodity problems, and there is no need to add auxiliary equipment, which greatly reduces maintenance and use costs.
project link
https://aistudio.baidu.com/projectdetail/6649316
In this project, Flying Paddle and Intel have established an industrial practice example of fresh food and retail product recognition systems, based on PP-ShiTuV2 Pipeline detailed model training, new varieties into the warehouse, and how to quickly deploy based on Intel OpenVINO™ to optimize the performance of CPU reasoning tasks, Extreme utilization of Intel x86 hardware resources.
scene difficulty
- Objects vary in shape, how to find the product to be detected?
- There are many varieties of commodities and fresh products, how to accurately identify the corresponding types?
- In the process of use, the iteration speed of commodities and fresh products is fast. How to reduce the cost of model update?
Design
As shown in the figure, for the above problems, we use the Pipeline in the figure to solve the above problems. The entire Pipeline is mainly divided into three parts:
subject detection
Detect commodities to be identified, remove redundant background information, and improve the accuracy of fresh product identification;
feature extraction
Extract features from images of fresh products to be identified;
retrieval module
Compare the features to be retrieved with the features of the fresh products in the database to obtain the labels of the fresh products to be retrieved.
Model Optimization Strategies and Effects
subject detection
Subject detection is a very widely used detection technology at present. It refers to detecting the coordinate position of one or more subjects in the picture, and then cropping the corresponding area in the image for recognition. Subject detection is the pre-order step of the recognition task. The input image is recognized after subject detection, which can filter complex backgrounds and effectively improve the recognition accuracy. Considering factors such as detection speed, model size, and detection accuracy, PaddleDetection's self-developed lightweight model PicoDet-LCNet_x2_5 was finally selected as the main body detection model of PP-ShiTuV2. The PicoDet series models integrate ATSS, Generalized Focal Loss, and cosine learning rate strategy , Cycle-EMA, lightweight detection head and other optimization algorithms. In addition, in order to better balance the detection speed and effect, PicoDet-LCNet_x2_5 replaced the CSP module in the neck with the LCNet module.
feature extraction
Feature extraction is a key part of image recognition. Its function is to convert the input image into a fixed-dimensional feature vector for subsequent vector retrieval. Considering the speed of the feature extraction model, model size, feature extraction performance and other factors, PaddleClas self-developed PPLCNetV2_base was finally selected as the feature extraction network. The PP-LCNetV2 model is optimized on the basis of PP-LCNetV1. It mainly uses the reparameterization strategy to combine the depth convolution of different sizes of convolution kernels, and optimizes point convolution, Shortcut, etc. Compared with PPLCNet_x2_5 used by PP-ShiTuV1, PPLCNetV2_base basically maintains a higher classification accuracy and saves 40% of inference time.
model deployment
Using OpenVINO ™ as the inference backend, the processing speed of tasks on the CPU side is multiplied.
The highly versatile x86 platform is used as the model deployment device, which can make full use of and realize multi-task loads, without the need to purchase additional accelerator card devices, which greatly saves project costs.
The final deployment environment of this project is: Intel x86 platform equipment. Considering the convenience of development, this example uses python to deploy the development environment. The system is a pipeline constructed by the three tasks of product detection, product feature extraction and product retrieval, and can realize a lightweight retail product recognition system based on a local data warehouse. A complete usage example and development instructions are also provided in the Xinghe community. You can refer to this tutorial to learn quickly, and develop and integrate for actual projects.
Wonderful course preview
In order to make it easier for friends to apply the example tutorials, OpenVINO™ software development engineer Ethan will give you an in-depth analysis of the entire development process from data preparation, scheme design to model optimization and deployment at 19:00 on August 31 (Thursday) . Teach everyone to practice code.
