Research On Deep Learning And Its Application On The Casting Defects Automatic Detection

Casting technology is widely used in automobile industry, aerospace fields because of its low cost and its ability of making complex parts. Propose a successful casting defects automatic detection algorithm has very important practical value because the algorithm plays an important role in improving the efficiency of the production of castings. Casting defects automatic detection has been a difficult and hot research in the field of pattern recognition due to the structural differences between different types of castings, the diversity of features of different kinds of defects, the randomness of location of defect area, etc. Convolution Neural Network(CNN) has ability to learn features from samples in that it simulates the mechanism of human’s processing flow of visual information, so CNN outperforms other deep learning model in the field of image processing. Regions with CNN(R-CNN) model which is based on CNN and its improved model —— Fast R-CNN and Faster R-CNN have a lot of improvement over the traditional targets detection algorithms.Casting defect detection algorithm based on Faster R-CNN model can solve the problem that casting defect area are hard to detect effectively, but in the mean time it’s hard to converge effectively in the processing of training. This paper studies the theory of deep learning theory, especially the Faster R-CNN model which is applied on the fileld of targets detection。By testing the quality of candidate boxes extracted by Region Proposal Networks(RPNs) and Selective Search, we mapped quality curve of candidate boxes, and pointed out the quality of candidate boxes extracted by RPNs is better than the quality of candidate boxes extracted by Selective Search. To solve the problem that the accuracy of small size defect area detected by Faster R-CNN model is not high, this paper put forward the anchor boxes set scheme, the scheme of the anchor boxes set can improve the accuracy of small size defect detected by Faster R-CNN model. For test samples set of casting defects in this paper, the experimental results show that Faster R-CNN model adopts the anchor boxes set scheme proposed by this paper spends more than 50 ms in detection, but improve the detection accuracy of small size air hole defects from 64% to 100%, the experimental results verify the validity of the scheme.Through comparison with the casting defect regions segmentation algorithm based on level set and casting defect areas classification algorithm based on CNNs, we prove that the Faster R-CNN model can obtain higher detection accuracy and faster speed, solve the problem of regional segmentation of casting defects.Finally, this paper implements the upper-computer control terminal software based on casting defects automatic detection algorithm proposed by this paper, and the software is applied in the actual production environment.