Digital instrument reading recognition based on improved SSD algorithm

A method for automatically reading digital instruments by improving the SSD algorithm is proposed to address the difficulties and time-consuming nature of manual reading. The backbone network of the SSD algorithm, VGG16, is replaced with an improved deep feature extraction network, ResNet50. A dilated convolution residual block is introduced using dilated convolutions to expand the receptive field size of ResNet50, allowing for more comprehensive extraction of features from small targets. A feature fusion module is constructed by progressively fusing effective feature layers of the SSD algorithm to address the issue of insufficient semantic and detail information in shallow feature maps. A feature attention module is introduced to enhance the algorithm’s ability to focus on important features. Experimental results on a self-made dataset of 1259 digital instrument images show that the improved SSD algorithm achieves an average precision mean improvement of 4.19% compared to the original SSD algorithm, with a reduction in computational complexity of 22.25 GFLOPs.