Pits are defects that occur during the film manufacturing process; they appear in the micrometer scale, which makes distinguishing them with the human eye difficult. Existing defect detectors have poor recognition rates for small objects or require a considerable amount of time. To resolve these problems, we propose a real-time small pit defect detector (RT-SPeeDet), a two-stage detection model based on an image processing and convolutional neural network (IP–CNN) approach. The proposed method predicts boundary boxes using a lightweight image-processing algorithm optimized for pit defects, and applies binary classification to the predicted regions; thus, simultaneously simplifying the problem and achieving real-time processing speed, unlike existing detection methods that rely on CNN-based detectors for both boundary box prediction and classification. RT-SPeeDet uses lightweight image processing operations to extract pit defect candidate region image patches from high-resolution images. These patches are then passed through a CNN-based binary classifier to detect small pit defects at a real-time processing speed of less than 0.5 s. In addition, we propose a multiple feature map synthesis method that enhances the features of pit defects, enabling efficient detection of faint pit defects, which are particularly difficult to detect.

No abstract