A dynamic commodity visual recognition method based on domain adaptation
LEI Yangyang, LI Li, SUN Fei, YAO Jian
2023, 0(3):
10-15.
doi:10.13474/j.cnki.11-2246.2023.0064
Asbtract
(
)
HTML
(
)
PDF (1972KB)
(
)
References |
Related Articles |
Metrics
Due to large deformation, occlusion, motion blur, similarity in appearance between items, and unknown distribution deviation in real scenes, item dynamic visual recognition still has huge challenges in practical applications. To this end, this paper proposes a dynamic commodity visual recognition method for smart retail. First, the bounding rectangle of the commodity is detected in real time through the target detection network, and then the category of the commodity is identified on this basis and recommendations are given to assist in the completion of consumer settlement. At the same time, in view of the cross-domain difference between the product picking video, the product library image, and the training image, this paper introduces a neighborhood style adaptive model (IBN) and a convolutional attention module (CBAM) to improve the domain adaptability of the model. In order to verify the effectiveness of this method, this paper constructs a real scene dataset Commodity247. The data is collected by the top-view camera of the smart container, including 247 common retail commodities and 37 050 pictures with annotated boxes and commodity categories. The experimental results show that on the Commodity247 dataset, the accuracy rate of product recognition(mAP) can reach 96.84%, the accuracy rate of the first recommendation(Rank1) can reach 98.41%, and the accuracy rate of the most difficult sample retrieval(mINP) can reach 85.24%, which is better than the one based on ResNet. For the basic model, mAP increases by 2.91%, Rank1 increases by 0.60%, and mINP increases by 10.86%, effectively reducing the influence of multi-angle, multi-light, and multi-background.