Research On Wine Foreign Body Detection Technology Based On Deep Learnin

Tiny impurities mixed in the bottled liquor will reduce product quality and damage customer’s health.Compared with the traditional manual detection,the detection method of tiny impurities based on machine vision has the advantages of high efficiency,high accuracy and low cost.The thesis designed a fast localization and detection algorithm for tiny impurities based on the deep learning object detection algorithm,applied it to detection of tiny foreign matter within bottled liquor,and made a data set about tiny foreign matter in bottle liquor.Specific research work is as follows:There is no publicly available dataset about foreign matters in bottled liquor,and the thesis first established an effective dataset.Many images of bottled liquor within tiny foreign matters were taken and more than 1500 sample sets were produced by enhancing and expanding original images with gamma transform,square transform,and CLAHE(Contrast Limited Adaptive Histogram Equalization).The thesis use the Faster R-CNN of object detection model to detect tiny foreign matters with low resolution and few effective features,which are difficult to be characterized by the top-level feature map in the bottled liquor and make improvements as follows: 1)The backbone network VGG16 is replaced with Res Net50 network to reduce the number of model parameters while maintaining a higher detection accuracy.2)The 7×7convolutional layers of Res Net50 is replaced with three 3×3 convolutional layers,and insert the CBAM attention module to improve the network’s ability to extract effective features from feature maps.3)With a combination of shallow features with strong position information and deep features,multi-scale detection is performed on multiple feature maps with different resolutions to reduce the pickup rate of tiny foreign matters.4)Replace the Ro I pooling layer with the Ro I Align network,retain the floating-point number boundary,and use bilinear interpolation instead of quantization to reduce the deviation between the suggestion box and the real label box.The improved model was tested on the self-made dataset.The results show that the Faster R-CNN object detection model based on the improved Res Net50-FPN can achieve more accurate detection of tiny foreign objects in bottled liquor.The average accuracy of the algorithm on the self-made dataset has reached 97.0%.