Publication: Detect and count food crackers from conveyer belt
Date
2024-07
Authors
Phang, Wei Jia
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Automation is a key to boosting productivity and efficiency in food manufacturing due to the fast-paced world of the food industry. The final goal of this project is to develop of a real time detection system as one of the automations to monitor the condition of the biscuit and the productivity of the production line using Deep Learning (DL). A convolutional neural network (CNN) is adopted in this work as an approach way including the data collection, pre-processing, and model training. CNN automatically learns features from image data, recognizing patterns at different levels of abstraction. This will increase the efficient and effective approach for real time object detections task. This report presents the development and evaluation of a YOLOv8 model for detecting and classifying 'Good' and 'Crack' biscuits. The dataset comprised 170 images including 80 ‘Good' and 90 'Crack' biscuits captured using an Oppo Reno4 smartphone. Data augmentation techniques such as rotation and translation were applied which expanding the dataset to enhance model training. The YOLOv8 model was trained and tested in a real-time detection system, leveraging a GPU-enabled environment for efficient processing. The performance of the model was evaluated using key metrics, including precision, recall, F1-Score, and mean average precision (mAP). The confusion matrix revealed that the model achieved perfect classification with no false positives or false negatives across various configurations, including YOLOv8x and YOLOv8n trained for 50 and 100 epochs. The YOLOv8x model demonstrated superior performance, achieving an accuracy of 92%, a precision of 93%, a recall of 90%, and an F1 score of 0.91 after 100 epochs, outperforming the YOLOv8n variant. Specifically, the model correctly classified 191 instances as 'Good' biscuits and 297 instances as 'Crack' biscuits. The precision, recall, and mAP metrics confirmed the model's robustness, demonstrating high accuracy and reliability for practical applications in quality control within the food industry. The prediction by uploading the image testing with both models have a high confidence score between the 97% to 99% in ‘Good’ detection and 96% to 97% in ‘Crack’ Detection. A Graphical User Interface was developed for the real time detection system. There is total 28 biscuits which is including 14 ‘Good’ biscuits and 14 ‘Crack’ biscuits are track and classified correctly by the real time system.