Developing the Smart Sorting Screw System Based on Deep Learning Approaches

Abstract

The deep learning technique has turned into a mature technique. In addition, many researchers have applied deep learning methods to classify products into defective categories. However, due to the limitations of the devices, the images from factories cannot be trained and inferenced in real-time. As a result, the AI technology could not be widely implemented in actual factory inspections. In this study, the proposed smart sorting screw system combines the internet of things technique and an anomaly network for detecting the defective region of the screw product. The proposed system has three prominent characteristics. First, the spiral screw images are stitched into a panoramic image to comprehensively detect the defective region that appears on the screw surface. Second, the anomaly network comprising of convolutional autoencoder (CAE) and adversarial autoencoder (AAE) networks is utilized to automatically recognize the defective areas in the absence of a defective-free image for model training. Third, the IoT technique is employed to upload the screw image to the cloud platform for model training and inference, in order to determine if the defective screw product is a pass or fail on the production line. The experimental results show that the image stitching method can precisely merge the spiral screw image to the panoramic image. Among these two anomaly models, the AAE network obtained the best maximum IOU of 0.41 and a maximum dice coefficient score of 0.59. The proposed system has the ability to automatically detect a defective screw image, which is helpful in reducing the flow of the defective products in order to enhance product quality.

A Novel Denoising Autoencoder Method for Surface Defect Detection of Screw Products

“…The mean square error is shown in equation (4), equation ( 5) and equation (6). Encoders and decoders are defined in equation ( 7) and equation (8). The schematic of the denoising autoencoder is shown in Figure 2.…”

“…It is not possible to implement supervised learning due to the requirement of a large number of defective samples for model training [2][3][4][5][6]. Unsupervised learning eliminates the weakness of supervised learning by only training normal images [7][8][9][10][11][12]. A study by Ke et al [13] used convolutional autoencoding techniques to detect and locate abnormal patterns in mobile phone logos.…”