In the practical applications of crack detection, traditional image processing based detection methods are faced with a lot of challenges, such as low level of intelligence, poor adaptability and so on. In this paper, an automatic pixel-level crack detection method based on deep transfer learning is proposed for engineering applications. The proposed detection method is composed of two stages: crack recognition and crack semantic segmentation, which makes it easy to meet the needs of efficient and reliable detection for large-scale collected images. The fine-tuned Vgg16 model in the first stage can accurately identify the crack images and avoid the computing cost for non-crack images in the further processing. The Unet++ model in the second stage provides a pixel-level semantic segmentation for crack images. Multifold knowledge based transfer learning and weighted loss function are proposed to improve the performance and generalization ability of the models. On 5 publicly available datasets, the proposed method achieved a mIoU of 84.62%. Tested on totally 8064 images with a resolution of 224×224, the proposed method achieved a mean detection speed of 8.1ms per image. Experiment results exhibit that the proposed method can effectively detect the pixel-level cracks. It provides an end-to-end deep learning method for large-scale crack detection for civil infrastructure. For the crack detection in complex scenes and the improvement of processing speed will be our next in-depth study.
In actual application, sensors are prone to failure because of harsh environments, battery drain, and sensor aging. Sensor fault location is an important step for follow-up sensor fault detection. In this paper, two new multi-level wavelet Shannon entropies (multi-level wavelet time Shannon entropy and multi-level wavelet time-energy Shannon entropy) are defined. They take full advantage of sensor fault frequency distribution and energy distribution across multi-subband in wavelet domain. Based on the multi-level wavelet Shannon entropy, a method is proposed for single sensor fault location. The method firstly uses a criterion of maximum energy-to-Shannon entropy ratio to select the appropriate wavelet base for signal analysis. Then multi-level wavelet time Shannon entropy and multi-level wavelet time-energy Shannon entropy are used to locate the fault. The method is validated using practical chemical gas concentration data from a gas sensor array. Compared with wavelet time Shannon entropy and wavelet energy Shannon entropy, the experimental results demonstrate that the proposed method can achieve accurate location of a single sensor fault and has good anti-noise ability. The proposed method is feasible and effective for single-sensor fault location.
Currently, for the task of salient object detection (SOD) based on deep learning, most approaches use a strategy of multi-level feature aggregation to enhance performance. However, due to the insufficient utilization of inter-pixel information, the aggregation of multi-level features often affects the prediction of salient objects and results in detecting blurry boundaries of salient objects. To tackle this problem, we have proposed a salient object detection network based on context-aware boundary perception. This network utilizes the context awareness (CA) branch to extract comprehensive contextual semantic information, guiding the network to focus attention not only on salient objects, but also on learning the mutual relationships between multiple salient objects. In addition, the boundary awareness (BA) branch is utilized to explore detailed boundary information around the contours of salient objects, enhancing the network's understanding of the edge pixels of salient objects. Moreover, we have introduced a new feature interaction aggregation (FIA) module, which is used to merge contextual semantic information and boundary detail information in the decoding stage to effectively utilize multi-level features and generate clearer and more accurate saliency maps. By conducting comprehensive experiments on three public datasets, we have demonstrated that our proposed method outperforms the current state-of-the-art representative methods.