A Hybrid CNN-Logistic Regression Model for Accurate Classification of Casting Defects

The abstract is an imperfect defect detection model meant to classify various defects of castings. It presents an excellent precision, recall, and $\mathbf{F 1}$-score of six classes of defects. For instance, the precision, recall, and F 1 -score for Blowholes are $97.98 \%, 97.70 \%$, and 97.84% respectively. Similarly, the Shrinkage Defects present with an impressive recall of $\mathbf{9 8. 0 8 \%}$ and an F1-score of $\mathbf{9 7. 9 5 \%}$. It can maintain high accuracy values for all of the classes concerned, showing 99% for them. The other inclusions seem to work similarly with $\mathbf{9 8. 0 3 \%}$. Cold Shuts, Hot Tears, and Metal Penetration contribute to this very robust capability in classification, exhibiting precision values such as $97.74 \%, 98.05 \%$, and 97.71% in their cases, respectively. In conclusion, there were $\mathbf{1 0, 6 5 5}$ images. The model achieves an average precision of 97.88% for all these images. The support of each class is fairly balanced; the proportions lie between 0.16 and 0.17, which suggests that each class is well represented in the dataset. This would ensure that there is no leaning of the performance of the model toward any one class, resulting in reliable and fair detection of casting defects. In addition, high support values are used to validate the accuracy of the model, and the detailed classification results validate the model’s ability to detect casting defects. Therefore, the proposed model proves to have outstanding performance by successfully detecting and classifying various defects in castings with 99% accuracy.