Commodity classification in livestreaming marketing based on a conv-transformer network

To obtain category information of goods in livestreaming marketing, we designed a recognition algorithm based on an improved transformer network that obtains the category of goods in live broadcast marketing via real-time detection of video images. First, based on the proposed parallel network structure, the local characteristics of the convolution module and global characteristics of the transformer module were merged, enriching the overall characteristics of the backbone network. Second, two different levels of global attention modules were used to calculate the feature map in groups to optimize the transfer module structure. The experimental results demonstrated that, under the same detector conditions, the detection accuracy and network parameters of the feature extraction network proposed in this paper was significantly improved. Furthermore, the detection effect on the commodity detection task in livestreaming marketing scene was significant.