Both transformer and one-stage detectors have shown promising object detection results and have attracted increasing attention. However, the developments in effective domain adaptive techniques in transformer and one-stage detectors still have not been widely used. In this paper, we investigate this issue and propose a novel improved You Only Look Once (YOLO) model based on a cross-attention strategy transformer, called CAST-YOLO. This detector is a Teacher–Student knowledge transfer-based detector. We design a transformer encoder layer (TE-Layer) and a convolutional block attention module (CBAM) to capture global and rich contextual information. Then, the detector implements cross-domain object detection through the knowledge distillation method. Specifically, we propose a cross-attention strategy transformer to align domain-invariant features between the source and target domains. This strategy consists of three transformers with shared weights, identified as the source branch, target branch, and cross branch. The feature alignment uses knowledge distillation, to address better knowledge transfer from the source domain to the target domain. The above strategy provides better robustness for a model with noisy input. Extensive experiments show that our method outperforms the existing methods in foggy weather adaptive detection, significantly improving the detection results.

For some measurement and detection applications based on video (sequence images), if the exposure time of camera is not suitable with the motion speed of the photographed target, fuzzy edges will be produced in the image, and some poor lighting condition will aggravate this edge blur phenomena. Especially, the existence of noise in industrial field environment makes the extraction of fuzzy edges become a more difficult problem when analyzing the posture of a high-speed moving target. Because noise and edge are always both the kind of high-frequency information, it is difficult to make trade-offs only by frequency bands. In this paper, a noise-tolerant edge detection method based on the correlation relationship between layers of wavelet transform coefficients is proposed. The goal of the paper is not to recover a clean image from a noisy observation, but to make a trade-off judgment for noise and edge signal directly according to the characteristics of wavelet transform coefficients, to realize the extraction of edge information from a noisy image directly. According to the wavelet coefficients tree and the Lipschitz exponent property of noise, the idea of neural network activation function is adopted to design the activation judgment method of wavelet coefficients. Then the significant wavelet coefficients can be retained. At the same time, the non-significant coefficients were removed according to the method of judgment of isolated coefficients. On the other hand, based on the design of Daubechies orthogonal compactly-supported wavelet filter, rational coefficients wavelet filters can be designed by increasing free variables. By reducing the vanishing moments of wavelet filters, more high-frequency information can be retained in the wavelet transform fields, which is benefit to the application of edge detection. For a noisy image of high-speed moving targets with fuzzy edges, by using the length 8-4 rational coefficients biorthogonal wavelet filters and the algorithm proposed in this paper, edge information could be detected clearly. Results of multiple groups of comparative experiments have shown that the edge detection effect of the proposed algorithm in this paper has the obvious superiority.

Machine learning-based defect identification has emerged as a promising solution to improving the defect accuracy of the aero-engine blade. This solution adopts machine learning classifiers to classify the types of defects. These classifiers are trained to use features collected in ultrasonic echo signals. However, the current studies show the potential number of features, such as statistic values, for identifying defect reaches a number more than that offered by an ultrasonic echo signal. This necessitates multiple acquisitions of echo signal and increases manual effort, and the feature obtained from feature selection is sensitive to the characteristic of the classifier, which further increases the uncertainty of the classifier result. This paper proposes an ensemble learning technique that is only based on few features obtained from an echo signal and still achieves a high accuracy of defect identification as that in traditional machine learning, eliminating the need for multiple acquisitions of the echo signal. To this end, we apply two well-known ensemble learning classifiers and simultaneously compare three widely used machine learning models on defect identification of blades. The result shows that the proposed ensemble learning models outperform machine learning-based models with an equal number of features. In addition, the two-feature-based ensemble learning model reaches an accuracy close to that of multiple statistic features-based machine learning models, where features are obtained from multiple collections of the signal.

Ultrasonic echo estimation has played an important role in industrial non-destructive testing and analysis. The ability to estimate parameters in the ultrasonic echo model is crucial to ensure the effectiveness of practical ultrasonic testing applications. In this paper, a scheme called ABIDE for identifying both multiple noises in the echo signal and the distribution of the denoised signal is proposed for ultrasonic echo signal parameter estimation. ABIDE integrates complementary ensemble empirical mode decomposition and the synchrosqueezed wavelet transform (CEEMD-SSWT) as well as the expectation maximization (EM) algorithm. The echo signal is split into a series of IMF components and a residual with the help of CEEMD, and then these IMFs are classified into the noise-dominant part and signal-dominant part by analyzing the correlation of each IMF and the echo signal using grey relational analysis. Considering the effect of noise in the signal-dominant part, SSWT is adopted to remove the noise in the signal-dominant part. Lastly, the signal output by the SSWT algorithm is used for reconstructing a denoised signal combined with the residual from CEEMD. Considering the distribution characteristic of the denoised signal, the EM algorithm is used to estimate parameters in the ultrasonic echo model. The relative performance of the proposed scheme was evaluated on synthetic data and real-world data and then compared with the state-of-the-art methods. Simulation results on synthetic data show that ABIDE outperforms the state-of-the-art methods in parameter estimation. Physical results on real-world data show that the proposed scheme has a greater PCC value in estimating echo model parameters. This paper also shows that ABIDE requires less convergence time than competitive methods.

The object detection task usually assumes that the training and test samples obey the same distribution, and this assumption is not valid in reality, therefore the study of cross-domain object detection is proposed. Compared with image classification, the cross-domain object detection task presents the greater challenge, which requires both accurate classification and localization of samples in the target domain. The teacher–student framework (the student model is supervised by pseudo-labels from the teacher model) has produced a large accuracy improvement in cross-domain object detection. Feature-level adversarial training is used in the student model, which allows features in the source and target domains to share a similar distribution. However, the direction and gradient of the weights can be divided into domain-specific and domain-invariant features, and the purpose of domain adaptive is to focus on the domain-invariant features while eliminating interference from the domain-specific features. Inspired by this, we propose a teacher–student framework named dual adaptive branch (DAB), which uses domain adversarial learning to address the domain distribution. Specifically, we ensure that the student model aligns domain-invariant features and suppresses domain-specific features in this process. We further validate our method based on multiple domains. The experimental results demonstrate that our proposed method significantly improves the performance of cross-domain object detection and achieves the competitive experimental results on common benchmarks.