Research On Defect Detection Of Injection Molded Parts Based On Faster R-CNN Algorithm Study

With the rise of a new generation of artificial intelligence and the introduction of the concept of Industry 4.0,enterprises including injection molding products have begun to transform from traditional models to intelligent and information-based digital factories.transformation provides strong support.In the transformation process of the injection molding factory to the smart factory,the automatic quality inspection of injection molding products is an indispensable part.However,at present,the quality inspection of injection molded products is carried out by the human eye,which is not only inefficient,but also cannot guarantee the stability of the product yield rate.Achieving automatic quality inspection of injection molded products can improve inspection efficiency,reduce labor costs,and improve production efficiency.With the development of industrial intelligence and automation,machine vision-based algorithms have been applied in defect detection of injection molded parts; however,because researchers are required to manually extract defect features,the robustness is poor and the versatility is lacking.Deep learning has the characteristics of automatic learning features and high prediction accuracy.It has developed rapidly in recent years and has been widely used in object classification,semantic segmentation,object detection and other fields.Therefore,this paper proposes to realize automatic defect detection of injection parts based on the Faster R-CNN model,and analyzes and further improves the Faster R-CNN model to achieve accurate detection of injection parts defects.The specific work of this paper is as follows:Preparation of injection molding data.When using deep learning models for training,the amount of data required is very large.Therefore,the data of the injection molded parts collected in practice is expanded by data enhancement techniques such as affine transformation,and the data of the injection molded parts pictures are marked by the Label IMG software,so as to prepare for the model training.Model improvements.Aiming at the problem of low detection accuracy and poor effect on small targets using the original Faster R-CNN model,and there are small targets such as bubbles in the injection molding data set in this paper,an improved method is proposed.A deeper Res Net50 residual network is used to replace the original VGG16 as the feature extraction network of the model; multi-scale feature fusion is introduced on the multi-layer feature layer for prediction,and the CBAM hybrid attention mechanism is added to the fused feature map.Further refinement; the newly generated feature maps integrate high-level semantic information and low-level defect information,and obtain richer feature information for small bubble sizes; ROI Align is used instead of ROI Pooling to improve small targets According to the analysis of the collected injection molding data set,the K-means++ method is used for clustering,and the size of the anchor box is reset according to the clustering results to improve the training speed and detection speed of the model.Experimental verification.In order to verify the effectiveness of the improved method for detection of small target defects in injection molded parts,the Faster R-CNN models before and after the improvement were trained on the injection molded parts training data set,and the detection effects of the models before and after were compared.From the results,it is found that the detection accuracy of the improved Faster R-CNN model has been improved,and the overall m AP has increased by 7.97%; especially the detection effect of small bubble target defects is more significant,and the AP value has increased from 76.29% to 99.42%,an increase of 23.13%.Online system design and implementation.Using the technology of separating front and back ends,a web-side injection molding defect detection system is designed.Researchers can pay attention to the status of the injection molding machine in real time by viewing the automatically detected defect information collected from the pictures of the injection molding machine,and can perform defect detection of injection molded parts by initiating a detection request on the browser side; in the function module,it is further verified that the model is aimed at small targets Validity of detection.