Abstract: There are problems such as cumbersome process, long processing time, and poor detection effect when the traditional image recognition algorithm is used to recognize the numbers in the instrument dial. Aiming at the above problems, a lightweight instrument dial detection method based on deep learning is proposed. Based on a single-shot multi-scale detection method (SSD), the proposal uses deep separable convolution instead of standard convolution to design feature extraction networks to improve feature expression capabilities and lightweight performance. At the same time, an anchor boxes construction process based on the distribution of ground truth boxes is proposed. In the process, an index that can quantify the matching degree of anchor frames (matching rate) is designed, and an anchor box scheme with a higher matching rate and a smaller number is constructed. Experimental results show that the proposed algorithm has less model parameters and computation, higher detection accuracy, and can obtain real-time processing speed in CPU environment.

Key words: lightweight feature extraction, anchor box design, smart meter detection, single shot multi-box detector(SSD)