An Efficient Method for DPM Code Localization Based on Depthwise Separable Convolution

With the popular application of direct part mark (DPM) technology, DPM code inspection has been a hot issue in the machine vision. It mainly consists of two steps, namely, localization and decoding. DPM code localization is a key and complex step in the DPM code inspection. However, the traditional localization methods suffer from complex imaging environment, involving various imaging background, illumination, imaging distance, and exposures. Furthermore, the target itself, i.e., the DPM code, could be severely polluted or worn. Aiming at improving the performance and robustness of DPM code localization, an efficient method with depthwise separable convolution is proposed in this paper. The optimized network model has the advantages of a few parameters, high computational efficiency, high precision localization, and good generalization ability. Meanwhile, the precision of the DPM code region is improved with the help of multi-scale prediction. The experiments on our DPM code localization database demonstrate the effectiveness and flexibility of the proposed method in comparison with the YOLOv3 network and the Tiny_YOLO network. Furthermore, the proposed method can estimate the exposure level of the DPM code region, which is benefiting to the DPM code recognition and enables the adaptive ability.