SOIF-DN: Preserving Small Object Information Flow With Improved Deep Learning Model

Object detection is a constantly evolving field of interest in the field of computer vision. As image-processing technologies continue to advance, detecting small objects has emerged as a critical challenge. However, currently available algorithms still struggle with this task owing to the small flaws in detection and complex backgrounds. This paper proposes a real-time detection method that specializes in recognizing small defects in images using an improved deep learning approach based on a feature pyramid network (FPN). The proposed method uses the small-object information flow (SOIF) module proposed by the backbone network to learn various features in images and quickly extract feature maps with varying gradient information. In addition, high-level feature maps are extracted from the FPN to maximize the information about small objects. These methods maximize the feature map properties of the existing layers and provide an opportunity to explore different features. The proposed deep learning model called SOIF-DN(SOIF-deep neural network) was evaluated on a public benchmark cell dataset and a private printed circuit boards (PCB) dataset. Upon reviewing the results, our method improves mAP@50 by 12.5% over YOLOv5 in terms of accuracy for small-object detection on PCBs. The speed performance (parameters/runtime) was 1.83 times better than YOLOv5 and 1.56 times better than YOLOv8. Additionally, our proposed SOIF-DN method for a different cell dataset enhances mAP@.5 by 3.8% over YOLOv5 for small object detection. The speed performance (parameters/runtime) was found to be 1.15 times better than YOLOv5 and 1.28 times better than YOLOv8.