Abstract: In order to improve the efficiency, accuracy and stability of packing counting in medical plastic packaging production line, this paper proposes a packing counting detection algorithm based on deep learning which can realize automatic online counting. Firstly, an improved RetinaNet network was constructed with ResNet as the framework, the feature pyramid network was used to generate multi-scale feature maps, and the convolution layers were cut appropriately. Then, clustering algorithm is used to optimize the Anchor size, so that the algorithm can adapt to count detection under abnormal conditions such as crooked bottle and inverted bottle, so as to reduce the missed detection rate and improve the positioning accuracy. Finally, the experimental evaluation of the algorithm on the actual packing data set shows that the algorithm is robust and reliable, and can quickly and accurately count and detect the packing plastic bottles under the production conditions. The counting accuracy can reach more than 99.98%, and the single detection time is 33 ms, which meets the real-time detection requirements of the production line.

Key words: deep learning, packing counting, RetinaNet, feature pyramid, clustering