Non-contact infrared detection is an effective way to diagnose substation equipment defects, but inclement weather conditions like haze and rainstorms produce low-quality infrared images, making it challenging to recognize equipment. To overcome this problem, a global and local contrast adaptive enhancement method is suggested. The global approach consists of the gray-level stretching Retinex (GLS Retinex) and the weighted truncated Gaussian histogram equalization (WTGHE). The GLS Retinex comprises gray-level stretching, fast-guide filtering, and parameter self-adjustment Retinex, which can be used to brighten dark images. The WTGHE contains Gaussian histogram calculation, truncated histogram equalization, and weighted RGB fusion, aiming to handle the blurred images. To significantly improve the quality of images in some detailed regions that cannot be handled by global enhancement algorithms, a local scheme called regional growth re-enhancement is proposed, which includes multi-seed region automatic selection and adaptive threshold growth criterion. The performance of the proposed technique is evaluated using over 1000 infrared thermal photographs of several substations under extremely difficult circumstances. The experimental results show that the novel strategy outperformed existing enhancement methods. The deep-learning detector also demonstrated the effectiveness of our proposed enhancement method for the improvement of detection precision.

A new method for 3D reconstruction that combines V-SLAM (visual simultaneous localization and mapping) and Cascade MVSNet (multi-view stereo network) is proposed in this paper, which enables the direct reconstruction of the 3D temperature field of substation equipment using video streams captured by a thermal infrared (TIR) imager. First, a V-SLAM-based pose estimation method is proposed to determine the position of the captured substation equipment. Parameter Self-Adjusted Retinex (PSAR) enhancement algorithm and Sparse Texture Feature Matching (STFM) method are developed to solve the problem of difficult extraction of TIR image features with weak texture. The key frame selection method based on deep learning object detection is proposed to improve the efficiency of the depth computation. Second, the improved Cascade MVSNet is used to compute the depth maps at the key frames in a hierarchical manner, and the loss functions of the scale and perspective consistency are utilized to eliminate the reconstruction errors resulted from different sizes and mutual occlusion of substation devices. The experimental results illustrate that this method is capable of directly reconstructing the 3D temperature field of a single substation device and a group of substation devices using TIR images, and the suggested algorithm is more advantageous in terms of accuracy, speed, and memory consumption compared with other state-of-the-art methods.

To address the challenges of unstable target localization and poor multi-module coordination in automated green pepper harvesting—caused by occlusions from branches and leaves, as well as varying lighting conditions—this paper presents the design and implementation of a modular robotic picking system. At the perception level, the system integrates a YOLOv8 detector with a RealSense D435i camera to identify and locate the calyx–ectocarp junctions of green peppers. A multi-target tracking with filtering algorithm is proposed, combining IoU-based association, Mahalanobis-distance-based matching, the Hungarian algorithm, Kalman filtering, and single exponential smoothing. This algorithm suppresses depth noise and trajectory jitter, thereby enhancing the stability and accuracy of 3D localization. At the control and execution level, a depth-first picking sequence strategy with ID freeze-state management is implemented within a multithreaded software–hardware co-design architecture. This approach avoids task conflicts and duplicate operations while supporting continuous multi-fruit harvesting. Field experiments under natural outdoor lighting and varying occlusion levels demon-strate that the proposed system achieves recognition rates of 91.57% and 80.29%, and harvesting success rates of 82.85% and 77.68% for non-occluded and lightly occluded fruits, respectively. The average picking cycle per pepper fruit is 9.8 s. This system provides an effective technical solution for addressing stability control challenges in the automated harvesting process of green peppers.

To address the challenges of unstable target localization and poor multi-module coordination in automated green pepper harvesting—caused by occlusions from branches and leaves, as well as varying lighting conditions—this paper presents the design and implementation of a modular robotic picking system. At the perception level, the system integrates a YOLOv8 detector with a RealSense D435i camera to identify and locate the calyx–ectocarp junctions of green peppers. An integrated multi-target tracking and filtering framework is proposed, which fuses multi-feature association, trajectory smoothing and coordinate denoising strategies to suppress depth noise and trajectory jitter, thereby enhancing the stability and accuracy of 3D localization. At the control and execution level, a depth-first picking sequence strategy with ID freeze-state management is implemented within a multithreaded software–hardware co-design architecture. This approach avoids task conflicts and duplicate operations while supporting continuous multi-fruit harvesting. Field experiments under natural outdoor lighting and varying occlusion levels demonstrate that the proposed system achieves recognition rates of 91.57% and 80.29% and harvesting success rates of 82.85% and 77.68% for non-occluded and lightly occluded fruits, respectively. The average picking cycle per pepper fruit is 9.8 s. This system provides an effective technical solution for addressing stability control challenges in the automated harvesting process of green peppers.

To address the challenges of rapid and accurate grape variety identification in natural orchard environments, along with the demand for efficient deployment on mobile devices, we propose in this paper YOLOv11n-GrapeLite, a lightweight model built upon an enhanced YOLOv11n architecture. First, an Efficient Channel Attention (ECA) mechanism is incorporated into the Neck layer. This mechanism adaptively recalibrates feature channel weights to emphasize those relevant to grape variety recognition, suppress background interference, and enhance target feature perception in complex scenes. Second, an adaptive downsampling (ADown) strategy is employed to replace the traditional convolutional downsampling module, reducing computational complexity while preserving critical features. Finally, the original C3k2 module is redesigned as a multi-scale convolution block (MSCB). This block integrates depthwise separable convolutions with multi-scale convolutions, which achieves significant parameter compression and enhances multi-scale feature extraction. Experimental results demonstrate that the proposed model achieves a mean average precision (mAP) of 91.5%, representing a 0.2% improvement over the original YOLOv11n, along with a 0.6% increase in recall. These results indicate outstanding robustness in complex field scenarios. The model’s parameter count was reduced to 1.87 million, computational complexity to 5.0 GFLOPS, and model size to 4.1 MB. These figures represent reductions of 27.8%, 23.1%, and 25.5%, respectively, compared to the original YOLOv11n, demonstrating significant lightweight optimization. Compared to mainstream models such as YOLOv6, YOLOv8n, YOLOv9s, YOLOV12, YOLOv13 and YOLOv26, the proposed model achieves superior performance in parameter count, computational load, and model size, while maintaining competitive detection accuracy. The YOLOv11n-GrapeLite model efficiently adapts to mobile terminal deployment, providing a feasible and efficient technical solution for real-time, precise identification of grape varieties in complex field scenarios.