Digital instrument identification based on block feature fusion SSD

Abstract

In order to identify digital instrument characters 0∼9 and decimal point in different scenarios, a digital instrument recognition algorithm based on block feature fusion SSD is proposed. Because the identification of small targets is difficult, in order to preserve the spatial information of small targets, the algorithm first divides the low-dimensional feature map into blocks and then fuses with the backbone network during the feature extraction phase. Secondly, in the prediction stage, the high-dimensional feature map is deconvoluted and then merged with the low-dimensional features to obtain the feature map with both spatial information and semantic information. Finally, the prediction result is passed through the character processing module to obtain the final representation. The experimental results show that compared with the original SSD, the algorithm improves the AP (Average Precision) of the decimal point by 30% and the mAP (Mean Average Precision) by 5.8%. It can accurately identify many different instrument representations in different environments and is robust enough to meet practical applications.