A convolutional neural network with equal-resolution enhancement and gradual attention of features for tiny target detection

Cheng, Minyang, Wang, Junliang, Zhou, Yaqin, Xu, Chuqiao, Liu, Ying ORCID: https://orcid.org/0000-0001-9319-5940 and Zhang, Jie 2022. A convolutional neural network with equal-resolution enhancement and gradual attention of features for tiny target detection. Presented at: IEEE 18th International Conference on Automation Science and Engineering (CASE 2022), Mexico City, Mexico, 20-24 August 2022. 2022 IEEE 18th International Conference on Automation Science and Engineering (CASE). IEEE, 10.1109/CASE49997.2022.9926425

Preview

PDF - Accepted Post-Print Version Download (586kB) | Preview

Abstract

The detection of tiny targets on the surface with high efficiency and accuracy is significant for the current intelligent manufacturing. Visual inspection methods based on deep learning are widely utilized to detect tiny objects. However, the tiny objects appear less distinct, less wide, and less area occupied in the image. At the same time, there is a lot of object-like noise, which further increases the difficulty of detecting tiny objects. In response to the challenges brought by the complexity of the detection environment, this paper proposes a detection network architecture that combines the enhancement of pixel-level features at equal resolution and the introduction of full-scale features based on attention. The model utilizes the subtle differences between the tiny target and the background and the semantic information of the tiny target outline to enhance the features of the tiny target while significantly reducing its loss in the equal-resolution feature layer. Additionally, a gradual attention mechanism is proposed to guide the network to pay attention to tiny objects features on the full-scale feature layer. The performance of this network architecture is validated on a real dataset. Experiments show that the model exhibits superior performance and outperforms existing resNet50, DenseNet, Racki-Net, and SegDecNet in detecting tiny objects.

Item Type:	Conference or Workshop Item (Paper)
Date Type:	Published Online
Status:	Published
Schools:	Schools > Engineering
Publisher:	IEEE
ISBN:	978-1-6654-9042-9
Date of First Compliant Deposit:	20 June 2022
Last Modified:	28 Nov 2022 16:22
URI:	https://orca.cardiff.ac.uk/id/eprint/150269

Actions (repository staff only)

Edit Item