| During industrial production,the surface of industrial components may have appearance defects such as scratches,roughness,dents,and other flaws due to reasons such as material,manufacturing process,collision,friction,etc.,and these defects can affect the service life of the parts.Due to the large structural differences in defect samples,industrial component defect segmentation is a highly challenging task.Currently,most deep learning-based industrial component defect detection algorithms do not fully utilize detailed features,resulting in segmented results lacking fine detail information.This information is crucial for the final accuracy of the network.For some networks,using only simple Convolutional Neural Networks(CNN)can capture local information and fine-grained features of the image well.However,the limited receptive field of CNN cannot capture global information and cannot dynamically adapt to changes in input.If the model lacks semantic understanding of the image,the final prediction result will be poor.For industrial components,the defect area is often small,and there are more backgrounds and normal areas,and there are few data samples,which presents a problem of small samples and small targets.We propose a semantic segmentation algorithm called In SeNet based on Transformer to address the aforementioned issues.It is a hybrid model consisting of a lightweight convolutional neural network Bi Se Net V2 and a lightweight and efficient Transformer.The powerful feature representation capability of Transformer is introduced into the segmentation algorithm,enabling the model to model long-range dependencies for defect segmentation.The improved Transformer structure is added to the fused detail and semantic branches of the Bi Se Net V2 network,which uses a specially designed efficient multi-head attention mechanism.Convolutional neural networks can utilize local feature information and have inductive bias such as translation invariance.By adding convolutional modules to Transformer,the model can better model local and global features of the image,improving segmentation performance.In addition,in order to enable the detail branch to capture multi-scale feature information without causing grid artifacts,we added mixed dilation convolutions after the detail branch to address these issues,and added a supervision structure to the detail branch to enable the model to better represent detailed features.The proposed model was evaluated on a self-annotated mobile device component dataset and a publicly available tile dataset,and achieved good segmentation performance. |
