DCWL-seg: PCB CT Image Element Segmentation Model Based on Dynamic Class Weighted Loss Function

One of the commonly used technologies in Printed Circuit Board (PCB) non-destructive testing is Computed Tomography (CT) technology, and element segmentation of CT images is a key subsequent step of non-destructive testing. Although the pretraining and finetuning paradigm greatly reduces the annotation workload of segmentation, the class imbalance problem caused by different pixel proportion of each class still seriously affects segmentation results of PCB CT images. For this reason, this paper proposes PCB CT image element segmentation model based on dynamic class weighted loss function (DCWL-seg). The model uses the baseline model EMLR in pretraining. In finetuning stage, feature fusion is added after feature transformation to improve the ability to utilize multi-layer features. Also, a loss function to alleviate class imbalance based on dynamic class weights (DCW Loss) is proposed. Dynamic class weights are set by IoU to guide model to focus on minority class pixels. Experiments show that mIoU of DCWL-seg on our PCB CT dataset reaches 89.8%, 1.2% higher than EMLR. And the IoU of small classes is more significantly improved, which proves the effectiveness of DCW Loss.