Dense Cut Tobacco Recognition Method Based on Improved DeepLabv3+

To solve the challenge of online detection of cut tobacco width during cut tobacco production, a lightweight dense cut tobacco segmentation method is based on an improved version of DeepLabv3+. An efficient channel attention mechanism module is introduced for dense cut tobacco in complex backgrounds to enhance the model's ability to extract tobacco features. Finally, the Atrous spatial pyramid pooling (ASPP) module is replaced by a cascade dilated convolution and receptive field fusion structure to improve the ability to capture global feature information to obtain rich cut tobacco features. The average pixel accuracy, average intersection-union ratio, and recall rate of the improved model were 84.34%, 68.2%, and 80.24%, respectively, 1.09, 2.61, and 1.99 percent points higher than the original model. The performance of the improved model was better than that of DeepLabv3+, Unet, and PSPNet models, and the processing time of a single image was 11.08 ms. This method realizes the segmentation of width measurement for cut tobacco, uncut scraps after cutting, and stuck strands in complex backgrounds, indicating that this research can provide valuable technical support for cut tobacco quality detection and online width measurement in the tobacco processing line.