Robotic vision is increasingly applied for surface inspection of built infrastructure. For this, it is essential to develop robust algorithms for semantic segmentation. This article presents a deep learning approach using a bidirectional self-rectifying network with Bayesian modeling (BSNBM) for improving detection accuracy, in dealing with the embedded uncertainty caused by false-positive labels and nonlinearity in sequentially convolutional blocks. For integration with residual encoders, a feature preserving branch is designed, wherein the output of previous dilated convolutional blocks is upsizedly or downsizedly passed on and concatenated with the following blocks recursively and bidirectionally. Further, to achieve robustness in feature representation with an acceptable level of credibility, convolutional kernels are randomized via a Bayesian model and adjusted per evidence update. As such, the network becomes less sensitive to uncertainty and redundant nonlinearity, which is inevitable in activation layers. Experimental results confirm the advantage of our BSNBM over current crack detection approaches.

In stereoscope-based minimally invasive surgeries (MISs), dense stereo matching plays an indispensable role in 3-D shape recovery, AR, VR, and navigation tasks. Although numerous deep neural network (DNN) approaches are proposed, the conventional prior-free approaches are still popular in the industry because of the lack of open-source annotated dataset and the limitation of the task-specific pretrained DNNs. Among the prior-free stereo matching algorithms, there is no successful real-time algorithm in none GPU environment for MIS. This article proposes the first CPU-level real-time prior-free stereo matching algorithm for general MIS tasks. We achieve an average$14-17$Hz on$640 \times 480$images with a single-core CPU (i5-9400) for surgical images. Meanwhile, it achieves slightly better accuracy than the popular efficient large-scale stereo matching (ELAS) method. The patch-based fast disparity searching algorithm is adopted for the rectified stereo images. A coarse-to-fine Bayesian probability and a spatial Gaussian mixed model were proposed to evaluate the patch probability at different scales. An optional probability density function estimation algorithm was adopted to quantify the prediction variance. Extensive experiments demonstrated the proposed method's capability to handle ambiguities introduced by the textureless surfaces and the photometric inconsistency from the non-Lambertian reflectance and dark illumination. The estimated probability managed to balance the confidences of the patches for stereo images at different scales. It has similar or higher accuracy and fewer outliers than the baseline ELAS in MIS, while it is 4–5 times faster.