Classification Detection of Packaging Bag Defects Based on Improved MobileNetV3

Aiming at the problems of high misdetection rate and low efficiency of the traditional packaging bag defect classification and detection task, this paper proposes a packaging bag defect classification and detection network based on improved MobileNetV3. Firstly, the Gaussian error linear unit (GELU) activation function, which has smooth nonlinear properties, is introduced into the Bneck block, to address the issue of model gradient vanishing. Secondly, the squeeze and excitation (SE) attention mechanism in the Bneck block is replaced by the coordinate attention (CA) mechanism, which considers channel information and direction-related position information. Thirdly, the position of the depth-wise convolution (DWConv) is adjusted, the optimal arrangement of convolutions in the Bneck block is investigated and determined. Ultimately, the experiments are conducted on the homemade packaging bag dataset, and the results demonstrate that the improved classification model reduces the intricacy of the model while achieving high accuracy in packaging bag defect detection.