Edible oil drums date detection based on boundary learning

In the industrial production environments such as irregular shapes of text, high reflectivity and blurring, the existing text detection method have the problems of inaccurate border localization and the limitation of detecting text within rectangular boxes. Based on boundary learning, a unified, broad-to-fine detection framework was proposed. It mainly consists of a feature extraction backbone network, a boundary suggestion module, and an iterative optimized boundary transformer module. ResNet network was used in the feature extraction backbone network to extract features from images. The boundary suggestion module consisted of multi-layer dilation convolution was used to generate rough bounding boxes. Additionally, an encoder-decoder structure was used in boundary transformer module to gradually improve the rough bounding boxes by iterative deformation under the guidance of classification map, distance field and direction field. The results of experiments conducted based on the self-collected datasets show that the accuracy, recall and F-measure values of the model are 91.56%, 87.38% and 89.41% respectively. The efficiency and advantages of the algorithm in the date detection of edible oil drums are verified.