Detecting and cleaning grain caking on the inner walls of silos is an important task to ensure food safety in storage facilities. However, in response to challenges such as insufficient lighting conditions, small and diverse forms of grain caking, this paper proposes the development and evaluation of a convolutional neural network model for robot vision detection of grain caking. The following improvements to the visual detection algorithm based on YOLOv5 are proposed in this article. Firstly, the Convolutional Block Attention Module (CBAM) and the improved Total Cross Union (CIoU) loss function are introduced to enhance the detection accuracy of grain caking. Secondly, by adding the Retinex Net algorithm with dark light enhancement, the recognition and detection performance under low light conditions can be improved. The improved YOLOv5 algorithm was trained and validated on a custom grain caking dataset. Comparative experiments show that compared with existing detection architectures, the improved algorithm has improved the average accuracy of grain caking detection by 1.8 % to 3.8 %. Finally, the improved algorithm proposed in this article was deployed on a wall climbing robot based on negative pressure adsorption, achieving real-time detection and automatic cleaning of grain caking.

We present a semiglobal stereo matching technique that combines enhanced census transform with unidirectional dynamic programming optimization. This method not only improves matching accuracy but also substantially reduces noise interference by using the 8-domain pixel median instead of the centre pixel in the standard technique and optimizing the initial matching generation value with the 8-path aggregation optimization algorithm. The experimental findings show that this suggested strategy efficiently avoids interference produced by various noises and reduces the false matching rate significantly.

Infrared thermography is a significant and extensively utilized method for assessing the operational condition of power equipment. Nonetheless, the constrained spatial resolution of infrared imaging systems, imaging noise, and the inadequate representational capacity of single-modality data render the precise segmentation of power equipment targets difficult, particularly in intricate backdrops and settings with weak structures. Simultaneously, obtaining high-quality pixel-level annotations for power equipment is expensive and laborious, leading to a scarcity of training samples and thus diminishing the efficacy of conventional supervised segmentation techniques. This research offers a super-resolution guided cross-modal segmentation strategy to tackle these issues in data-scarce circumstances and examines the applicability of the general-purpose segmentation model Segment Anything Model 3 (SAM3) for infrared image segmentation of power equipment. A super-resolution reconstruction framework based on a high-order degradation model is built to enhance low-resolution infrared images collected in real-world contexts. An Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN) -based network incorporating residual-in-residual dense blocks (RRDB) is utilized to reconstruct infrared thermograms, hence improving structural features and boundary representations. Secondly, the concurrently obtained visible-light images are improved by low-light enhancement methods, and an anchor-free object detection framework is employed to ensure accurate localization of power equipment targets. The identified areas in visible images are aligned with the coordinate system of infrared super-resolution images via cross-modal geometric transformation, establishing a cross-modal spatial prior that efficiently limits the search space for infrared segmentation and mitigates background interference. The general-purpose segmentation model SAM3 is introduced, utilizing cross-modal detection boxes as prompts to facilitate precise segmentation of power equipment targets in infrared super-resolution images, achieving high-accuracy segmentation without the necessity for extensive task-specific annotated data. The experimental results demonstrate that our proposed approach significantly improves both the accuracy and robustness of infrared image segmentation for power equipment under complex conditions, attaining a Jaccard index of 89.86% and a Dice coefficient of 91.12%, thereby validating its efficacy and practical applicability in data-scarce environments.