Enhanced Cigarette Pack Counting via Image Enhancement Techniques and Advanced SAFECount Methodology
Yanghua Gao, Zhenzhen Xu, Xue Xu · Traitement du signal · 2023
In the realm of cigarette pack counting systems, prevalent challenges persist, notably the low accuracy in count, limited adaptability to intricate scenes and varying environments, and a lack of responsiveness to diverse pack types and shapes.This study introduces an advanced method for cigarette pack counting, leveraging a combination of various image enhancement techniques and an improved Similarity-Aware Feature Enhancement block for object Counting (SAFECount) approach.The methodology comprises three integral modules: an image enhancement module, a feature extraction module, and a counting module.The image enhancement module, tasked with noise reduction and deblurring, ensures targeted enhancement effects on cigarette box images.To contend with the rapid shifts in cigarette pack appearances, this research integrates specialized color and boundary feature extraction networks with the SAFECount method.This integration facilitates the fusion of multi-scale, key semantic information, thus amplifying the model's detection efficacy.Addressing the scalability limitations prevalent in general models, the study employs a few-shot counting (FSC) approach, which endows the model with essential generalization and flexibility, requisite for practical applications, even with a minimal training dataset.Empirical analyses, conducted using actual data from the Zhongyan Corporation's cigarette pack dataset, substantiate the superiority of the proposed method in real-world warehouse environments.The method demonstrates a marked improvement in counting performance, evidenced by a Mean Absolute Error (MAE) of 1.71 and a Root Mean Square Error (RMSE) of 1.95.