| dc.description.abstract | Tronrud engineering’s packaging machines have occurrences of lost products that exceed the expectations of their customers, and there is currently no optimal way to detect it. The goal of this thesis is to use computer vision to solve this problem by developing two different computer vision models. Where Model 1 will use image comparison to evaluate the similarity between an input image and a target image to detect product loss. Model 2 are using a convolutional neural network to extract features from an image. The features extracted in the network are then used to distinguish if there are any errors or anomalies in the product in the image. The implementation includes data collection and methods to preprocess the data such as grayscaling, thresholding, cropping, and rotating before the models use this data. With the data collected, both Model 1 and Model 2 were able to get 100% prediction accuracy on the output from the test data. However, for Model 2, the data collected was limited and the model should be trained and tested on a bigger dataset. Deployment of the models connected to live data from the machines was a shortcoming in this thesis, but suggested solutions for deployment and implementation are given for further work. | |
