On the ADE20K international list of scene parsing and semantic segmentation technologies released in March 2022, MYbank ranked first with a new high of 0.6730, which is another authoritative list in the field of computer vision after winning and maintaining the cityscapes semantic segmentation list last year.
The ADE20K is a dataset published by the Massachusetts Institute of Technology (MIT) in 2017 that can be used for a variety of tasks such as scene perception, segmentation, and multi-object recognition, containing more than 20,000 photos of more than 150 different indoor and outdoor scenes, and participants need to accurately segment and mark objects such as buildings, trees, people, glass windows, etc. through algorithms, and the higher the accuracy, the higher the score. Compared with the same type of cityscapes, the environment and scenario of ADE 20K setting are more complex and the challenges to algorithm adaptability are greater, and it is the authoritative benchmark dataset for semantic segmentation papers of the world's three top conferences (CVPR, ICCV and ECCV).
In view of the characteristics of the ADE20K dataset, the online provider proposed the BKSeg algorithm, and after optimizing the network structure and improving the training strategy, the Pixel Accuracy reached 0.8158, the mIoU reached 0.5301, and the final score was 0.6730, ranking first.
The BKSeg algorithm is optimized based on the Mask2Former algorithm framework. Specifically, the backbone neural network based on the combination of CNN and Transformer is first used for feature learning; secondly, the Multi-scale FAPN strategy is combined to improve the segmentation effect of objects of different sizes; in addition, the BKSeg algorithm also adds OHEM module to improve the segmentation effect of difficult to identify objects; finally, the model effect is further optimized in combination with Semi-Supervised Learning technology
It is understood that MYbank is currently the only institution that has maintained a leading position on both international lists, and there are also global technology companies and research institutions such as Huawei, SenseTime, Microsoft, Google, and Amazon.
Scene parsing and semantic segmentation are important technologies of computer vision, which can help computers accurately perceive and understand objects on flat photos, thereby transforming unstructured images into data that can be calculated, providing a basis for restoring and understanding real-world scenes. The objects in most scenes are diverse and complex, and the same type of objects in different scenes have different colors, shapes, sizes, and postures, so how to segment and accurately identify different objects through optimization algorithms is the difficulty of such technologies.
The continuous development of computer vision has also brought new methods and models to financial services. As a technology bank focused on serving small and micro enterprises initiated by Ant Group, MYbank has widely applied such technologies in the fields of rural finance and small and micro financial risk control.
In the field of rural finance, MYbank took the lead in introducing satellite remote sensing image recognition technology, and carried out 10X10 meters pixel-level crop recognition through computer vision on medium and low-resolution satellite images, realizing the recognition of more than 20 kinds of staple grain crops such as corn, rice, wheat and more than ten kinds of cash crops such as apples, kiwis and citrus, and achieved 93% recognition accuracy that could only be achieved with high-resolution satellite images in the past. Based on this technology, the MYbank's Big satellite remote sensing risk control system has now covered more than 1,000 counties across the country and served hundreds of thousands of large growers.
At the same time, through the continuous optimization of the visual image recognition segmentation algorithm, it is also possible to automatically extract facilities such as farmhouses and greenhouses, and quickly identify information such as the type, quantity, and area of facilities, so as to obtain regional distribution information of rural economic activities and guide the credit level to provide better regional services.
As shown in the following figure, the basemap is a public image of the sky map area, with red (farmhouse) and green (greenhouse) marking the identified facilities respectively.
In addition, in the scene of the offline physical store, image perception technology can also be used to identify, understand and portray the industry and business behavior of the store. For example, by identifying the store front photos, shelves and even purchase and sale of contracts and invoices taken and uploaded by users, and combined with multi-dimensional cross-verification technology, online business banks can more accurately portray the industry, upstream and downstream relationships and real business appearance of small and micro operators, thus providing a basis for operating loan credit.
Legend: By analyzing the photos, you can accurately identify the goods on the shelves of the supermarket, thus providing a reference for analyzing the sales situation