23 August 2023 in IEEE Transactions on Instrumentation and Measurement
Deep learning-based surface defect detectors play a crucial role in ensuring product quality during inspection processes. However, accurately and efficiently detecting defects remains challenging due to specific characteristics inherent in defective images, involving a high degree of foreground-background similarity, scale variation, and shape variation. To address this challenge, we propose an efficient Transformer-based detection network, ETDNet, consisting of three novel designs to achieve superior performance. Firstly, ETDNet takes a lightweight Vision Transformer to extract representative global features. This approach ensures an accurate feature characterization of defects even with similar backgrounds. Secondly, a channel-modulated feature pyramid network (CM-FPN) is devised to fuse multi-level features and maintain critical information from corresponding levels. Lastly, a novel task-oriented decoupled (TOD) head is introduced to tackle inconsistent representation between classification and regression tasks. The TOD head employs a local feature representation module to learn object-aware local features and introduces a global feature representation module, based on the attention mechanism, to learn content-aware global features. By integrating these two modules into the head, ETDNet can effectively classify and perceive defects with varying shapes and scales. Extensive experiments on various defect detection datasets demonstrate the effectiveness of the proposed ETDNet. For instance, it achieves AP 46.7% (v.s. 45.9%), and AP 50 80.2% (v.s. 79.1%) with 49 FPS on NeU-DET. Code is available at https://github.com/zht8506/ETDNet.
21 June 2024 in IEEE Transactions on Neural Networks and Learning Systems
The detection head constitutes a pivotal component within object detectors, tasked with executing both classification and localization functions. Regrettably, the commonly used parallel head often lacks omni perceptual capabilities, such as deformation perception (DP), global perception (GP), and cross-task perception (CTP). Despite numerous methods attempting to enhance these abilities from a single aspect, achieving a comprehensive and unified solution remains a significant challenge. In response to this challenge, we develop an innovative detection head, termed UniHead, to unify three perceptual abilities simultaneously. More precisely, our approach: 1) introduces DP, enabling the model to adaptively sample object features; 2) proposes a dual-axial aggregation transformer (DAT) to adeptly model long-range dependencies, thereby achieving GP; and 3) devises a cross-task interaction transformer (CIT) that facilitates interaction between the classification and localization branches, thus aligning the two tasks. As a plug-and-play method, the proposed UniHead can be conveniently integrated with existing detectors. Extensive experiments on the COCO dataset demonstrate that our UniHead can bring significant improvements to many detectors. For instance, the UniHead can obtain $+$ 2.7 AP gains in RetinaNet, $+$ 2.9 AP gains in FreeAnchor, and $+$ 2.1 AP gains in GFL. The code is available at https://github.com/zht8506/UniHead.
18 October 2022 in Image and Vision Computing