A multiwindow detection method for tobacco impurities in images based on multilayer convolutional networks

Current methods for detecting impurities in tobacco images rely on extracting single features, leading to low detection accuracy. Therefore, this paper studies a multi-window detection method for tobacco impurities based on multi-layer convolutional networks. Preprocessing of the tobacco impurity images improves image quality. Due to the large size differences of impurities, a multi-scale window segmentation strategy is adopted to comprehensively cover impurity targets of different sizes. The image is input into a multi-layer convolutional network, which automatically learns multiple features of the tobacco impurity image using convolution and pooling layers. A feature pyramid network is introduced to fuse multi-scale features, and the classification result is output through a fully connected layer. The network is trained with a large amount of labeled data, and the detection results of impurities in the tobacco images are finally obtained and visualized. Experimental results show that the multi-window detection method of multi-layer convolutional networks can accurately detect impurities in tobacco images. The specificity reaches 0.987 with 800 samples, which is higher than the comparison method. The ROC curve shows that the TPR reaches 0.97 with an FPR of 0.5, and the curve remains optimal. This method provides an excellent technical solution for tobacco quality inspection.