Abstract: To solve the problem of a small number and variety of defect samples in defect detection, which makes it difficult to provide enough data for training of supervised deep learning models, this paper uses a large amount of easily accessible positive sample data （without defects） in industrial production to built an auto-encoder model of convolutional network with Encoder-Decoder structure, and embeds the convolutional attention module combining spatial and channel attention in the encoder to enhance the network feature extraction ability. We added the contextual information module in the encoding stage to obtain a larger perceptual field and reduce the computational requirements. Meanwhile, multi-scale structural similarity MS-SSIM and L1 loss were combined to improve the quality of reconstructed images, and peak signal-to-noise ratio （PSNR） was used to measure reconstruction error and discriminate anomalies. The experimental results show that the proposed medical glass bottle defect detection method can accurately detect the defect data and segment the defect region with 99.45% accuracy, 97.63% recall rate, 0.55% miss detection rate, and 2.93% false detection rate. The method can accurately detect the glass bottle defect and locate the defect area, and the image reconstruction time is short, only about 10.37 ms, which can achieve accurate and efficient automated product quality inspection.

Key words: medical glass bottles, defect detection, convolutional self-encoder, attentional mechanism, peak signal-to-noise ratio