An Attention-Based Convolutional Neural Network With Spatial Transformer Module for Automated Optical Inspection of Small Objects

In the manufacturing industry, automated optical inspection (AOI) systems play a crucial role in accurate and efficient inspection. Recently, convolutional neural networks (CNNs) have become increasingly popular for this AOI because they can automatically extract optimal features through convolutional and pooling operations. In this article, we aim to classify product images obtained from an AOI system based on the product codes in them. However, since the size of the product codes is very small compared with the image, it is challenging to achieve sufficient classification accuracy using a standard CNN architecture. To address this issue, we propose an attention-based CNN (A-CNN) model that focuses on small objects within images. The A-CNN model integrates an attention module to adaptively extract regions of interests (ROIs) centered on small objects, and a classification network (CN) to classify these ROIs, thereby increasing the object-to-image area ratio (OAR) and improving classification accuracy for small objects. The A-CNN model can be effectively trained end-to-end with minimal data labeling compared with object detection methods. Experimental results show that the proposed A-CNN model achieves a classification accuracy of 99.92% and an inference speed of 33 frames/s on the NVIDIA Jetson Nano platform, outperforming the smallest models of YOLOv5, YOLOv7, YOLOv8, YOLOv9, and YOLOv10, state-of-the-art object detection algorithms, in terms of both accuracy and latency. Notably, our model has $2.0\times $ faster than the fastest YOLO model, underscoring its efficiency in real-time applications. These findings highlight the potential of the A-CNN model as an accurate and practical solution for small object classification.