Efficient Neural Net Approaches in Metal Casting Defect Detection

One of the most pressing challenges prevalent in steel manufacturing industry is identification surface defects. Early casting defects can help boost performance, including streamlining production processes. Though, deep learning models have helped bridge this gap and automate these processes, there a dire need to come up with lightweight that be deployed easily faster inference times. This research proposes architecture efficient terms accuracy time compared sophisticated pre-trained CNN architectures like MobileNet, Inception, ResNet, vision transformers. Methodologies minimize computational requirements such as depth-wise separable convolution global average pooling (GAP) layer, techniques improve architectural efficiencies augmentations, been experimented. Our results indicate custom model 590K parameters convolutions outperformed pretrained Resnet Vision transformers (81.87%) comfortably outdid Resnet, times (12 ms). Blurpool fared other techniques, an 83.98%. Augmentations had paradoxical effect on performance. No direct correlation between 3 × time, they, however, they played role improving efficiency by enabling networks go deeper decreasing number trainable parameters. work sheds light fact built without relying architectures.