LCCD-Net: A Lightweight Multiscale Feature Fusion Network for Cigarette Capsules Defect Detection

Defect detection of cigarette capsules is crucial for maintaining product quality in tobacco manufacturing. The intricacies in the cigarette capsule process present several challenges to address: 1) detection of small-scale defects, 2) complex defect features, and 3) lightweight requirements for industrial applications. In this article, we propose a lightweight cigarette capsules detection network (LCCD-Net) to overcome the crucial challenges. Concretely, for small-scale defects, we introduce a Star Operation (element-wise multiplication) Mixed Aggregation Network (S-MANet) that implicitly maps features to a high-dimensional, nonlinear feature space. Based on S-MANet, we construct a neck structure called HyperStarNeck to enhance feature fusion and improve small-scale feature capture. Meanwhile, to capture complex features, we design a Dynamic Multi-scale Cross Stage Partial (DMS-CSP) module that dynamically generates convolution kernels based on input multi-scale features. Moreover, we propose a Separated and Enhancement Attention Detection Head (SEAD-Head) that reduces the interference of complex backgrounds by improving feature map attention allocation, while lightening the structure for deployment. On the real-world Cigarette Capsules Bubble and Tail Defect Dataset (CCBTD), our approach achieves 85.3% average precision (AP) with only 1.04M parameters and 5.02 GFLOPs, surpassing state-of-the-art methods. Furthermore, evaluation on our Synthetic Industrial CCBTD (SI-CCBTD) dataset, which simulates diverse industrial environments, yields 81.1% AP, demonstrating robust performance under varying conditions. Additionally, our method attains 62.0% AP on the synthetic aperture radar ship detection dataset (SSDD), validating its cross-domain effectiveness. Our source code and SI-CCBTD dataset are available at IEEE DataPort, doi: 10.21227/snr3-rj68.