Research On Casting Appearance Defect Detection Based On Deep Learning

The surfaces of castings often have defects like dark hole,shallow pits,crack,bulge and depression.These defects not only weaken the strength of the casting but also cause structural deformation,which lead to a safety hazard.At present,traditional detection methods have been difficult to meet industrial demands due to high cost,low efficiency,and strong artificial dependence.In this study,we proposes a depth learning based casting defect detection algorithm to achieve better detection results.In order to preserve the defect features as much as possible,we use a 5 megapixel industrial camera to capture images of 2566*1940 pixels.The algorithm is mainly divided into two parts:Firstly,we train a YOLO network to find the area to be detected from the original image,segment the area image and adjust it to 256*256 pixels.We then used an improved ResNet-50 network to identify the resulting image and determine if it was defective.In order to improve the detection accuracy,we made two improvements to the ResNet-50 network: 1.We proposed the ASoftReLU function to replace the traditional ReLU activation function,which avoids neuronal death and further improves the training speed and classification accuracy of the network.2.We designed a multi-channel convolutional neural network to combine data expansion with image feature enhancement to consolidate the essential features of the image during training and enhance the anti-jamming capability of the network.To verify the effectiveness of the algorithm,three sets of experiments were carried out on the TensorFlow platform to build a neural network.The last set of experiments was the final result of the algorithm.The results of the experiment have two judgment indicators: 1.Whether there is a defect;2.The category of the defect.According to the experimental results,the improved algorithm achieves 98.2% training set accuracy and 94.3% test set accuracy in the task of determining whether there is a defect.For the determination of defect categories,the algorithm achieved 88.2% accuracy on the test set.Both tasks achieved good accuracy,indicating that the improved network has good detection ability.