The work aims to extract the cigarette pack image data to train the deep learning target detection model, and improve the efficiency and accuracy of cigarette pack assembly lines. A cigarette pack recognition and classification model was established based on deep learning to improve the original YOLOv3 model. The designed multi-space pyramid pooling structure (M-SPP) was added to the original network. The downsampling of the 64×64 feature map was spliced with that of the 32×32 feature map. The prediction feature layer of 16×16 was removed to improve the detection accuracy and speed of the model, and the K-means++ algorithm was used to optimize the a priori frame parameters. The experiment showed that the average accuracy of the target detection model reached 99.68%, and the detection speed reached 70.82 frames per second. It is concluded that the image recognition and classification model established based on deep learning has high accuracy and fast detection speed, which can effectively meet the automatic real-time detection of cigarette pack assembly lines.