An Adaptive FD-Former Network For Tobacco Impurity Detection

In the tobacco manufacturing industry, tobacco impurity detection. However, the complicated background settings and the morphological diversity of impurities during production can compromise the detection accuracy of tobacco robots, thereby affecting product quality. To address these challenges, we present a novel online detection method for tobacco impurities designed for tobacco inspection robots. We introduce the DAC-Former attention mechanism module, which effectively minimizes the impact of extraneous image information. Furthermore, we propose the Double Feature Fusion Network to boost detection accuracy by merging deep semantic and surface location data from feature maps. A dedicated dataset for tobacco impurity detection has been created to validate our approach. Experimental results indicate that our algorithm enhances the mAP value by approximately 8.71% and efficiently tackles tobacco impurity detection by enabling real-time model processing at a higher frame rate compared to the conventional YOLOX method.