Abstract:

Since the data sets of foreign objects in the conveyor belt collected by the video detection algorithm based on deep learning for foreign objects in the conveyor belt were fewer, this paper proposed a detection algorithm based on YOLOX for foreign objects in the conveyor belt. First, MobileNet V3 was used to replace the backbone feature extraction network CSPDarknet in YOLOX to reduce the calculation time of the backbone features, and then the convolutional network in the enhanced feature extraction network in YOLOX was replaced with a deep separable convolutional neural network to reduce the calculation time of the enhanced feature extraction network. Finally, a convolutional block attention module (CBAM) was added before the deep separable convolutional neural network in the enhanced feature extraction network to improve detection accuracy. Experiments have verified that the algorithm used in this paper can complete the training and achieve a recognition accuracy of 96.5% when using a small number of samples. At the same time, the average time of this algorithm is only 0.019 s, and the recognition frame rate can reach 40 FPS.

Key words: small sample, foreign object detection, transfer learning, YOLOX, convolutional block attention module (CBAM)