Detection of cigarette appearance defects based on improved SSD model

作者：Rui Qu, Guowu Yuan, Jianchen Liu, Hao Zhou · 发表于：Proceedings of the 2021 5th International Conference on Electronic Information Technology and Computer Engineering · 年份：2021 · DOI：10.1145/3501409.3501612 · 被引用次数：9 · 研究领域：Industrial Vision Systems and Defect Detection、Advanced Neural Network Applications、Vehicle License Plate Recognition

The automatic detection of product defects has been applied to the assembly line production of various industries. With the automation of the production of cigarettes, the detection of appearance defects of cigarettes can no longer be completed manually. Aiming at the current low efficiency of cigarette appearance defect detection, this paper proposes a model based on improved SSD network and pyramid convolution for cigarette appearance defect detection. First, replace the original VGG16 feature extraction network with the ResNet50 network, which has better feature expression capabilities, to improve the performance of small target detection while retaining more shallow semantics; Secondly, replace the original 3×3 convolution in the network and use pyramid convolution to expand the receptive field and extract multi-scale information; At the same time, the original loss function is optimized to solve the problem of unbalanced cigarette appearance defect categories in the data set. The experimental results show that the average accuracy of the algorithm in this paper is 90.26%, which is much higher than the accuracy of the original SSD model and other target detection algorithms.