当前工业化养殖胡瓜钝绥螨的产量巨大,导致现有人工品质管控方式无法及时准确地对螨虫养殖品质进行监测.设计研发一款螨虫监控系统,包含螨虫麦麸分离装置、显微视觉装置、改进的YOLOv5 螨虫检测模型.通过对YOLOv5 模型的网络结构进行分析,在原YOLOv5 模型上将嵌入ECANet注意力机制的MobileNetv3-Large网络作为主干特征提取网络,加强网络提取特征能力,进一步通过裁剪模型预测头的冗余部分,提升模型对中小目标检测性能.实验结果表明:改进后的M3ECA-YOLOv5-2H模型对比原YOLOv5 模型,在平均检测精度mAP50 和mAP75上提高了 0.68 和 4.62 个百分点;在模型大小上降低 18.9 MB,占用更低内存;在单张图片推理速度上提升 4 ms.此外对比其他 5 种通用目标检测模型,M3ECA-YOLOv5-2H模型在检测精度和检测速度等指标上整体表现更佳,具有很好的应用价值.

牙模 3D打印过程中出现的质量缺陷影响着牙模的外观和使用性能,严重时导致废品率高,造成时间、材料的浪费.为了减少牙模 3D打印的废品率,提出了一种基于YOLOv5 的牙模 3D打印实时缺陷检测方案:首先获取多维度牙模缺陷图片,采用切图分割与数据增强处理等方式制作数据集,然后通过构建YOLOv5 深度学习网络模型对牙模 3D打印图像进行迭代训练,最后通过YOLOv5 程序调用摄像头并使用训练后最佳权重值进行实时在线缺陷检测.通过实验对比,YOLOv5 深度学习网络模型的检测准确率要高于Faster R-CNN、YOLOv3、SSD这 3 种模型,其检测平均准确度高达 94.78%,平均检测时间为 21 ms.结果表明该方法能够检测牙模 3D打印过程中的质量缺陷问题.

针对胡瓜钝绥螨体积小、与粉螨相似、难以分类计数问题,提出一种基于YOLOv4胡瓜钝绥螨自动检测计数方案.在多时间段、分批次拍摄螨虫照片制作数据集,通过构建YOLOv4深度学习模型对螨虫图像进行特征提取和高精度分类检测.结果表明,基于YOLOv4的螨虫识别模型能在图像存在杂质且螨虫相连情况下准确检测胡瓜钝绥螨,平均检测单幅图片耗时约0.137 s.在测试集中对胡瓜钝绥螨检测精度达到99.45％,粉螨检测精度达到93.94％.对胡瓜钝绥螨和粉螨检测精度均值(mAP)达到96.69％,分别比YOLOv4-Tiny、YOLOv3、Faster-RCNN、EfficientDet模型高14.85、2.12、35.77、27.18百分点.设计螨虫GUI(图形用户界面)品质监控界面,方便检测人员操作.本研究提出的检测方案具有很高的应用价值,能够精准、快速识别胡瓜钝绥螨,可为胡瓜钝绥螨养殖质量检测提供技术支撑.

With the continuous change of artificial intelligence technology, the study of path planning for mobile robots is no longer limited to traditional path algorithms. Reinforcement learning, as an artificial intelligence algorithm with excellent performance in the field of path planning, has also gradually become the object of research on path planning. In order to find a fast path from the starting point to the end point of a mobile robot in a complex environment, reinforcement learning is used to find a valid path by sensing the environment and continuously receiving rewarding feedback through a method of trial and error learning like humans. Therefore, this paper selects two algorithms to verify their effectiveness on the basis of building a two-dimensional grid map. The final experimental results show that both reinforcement learning algorithms can eventually avoid obstacles and plan a valid path in a complex environment through continuous learning iterations. Q-Learning also reduces the path distance by 10% and the number of convergence iterations by 82% compared to Sarsa.

The problem of birds nesting in high-voltage towers has laid a major hidden danger to the safe operation of long-distance transmission lines. In the power line inspection, the image background of the bird’s nest is complex. Some nests are small and locally occluded, and it is difficult for existing object detection algorithms to detect with low computational effort under high accuracy. To solve the above problems, we recommend using KBN-YOLOv5. Based on YOLOv5, we use K-means algorithm to cluster the size of bird nest image to set the size of anchor frame. In the Backbone network, we replace the Focus and SPP modules with Conv and SPPF modules, respectively, and adjust the number of BottleneckCSP modules as well as their positions. Finally, in the Neck network, we improve the BottleneckCSP module combined with the Efficient Channel Attention (ECA) module for a more effective weight information distribution to accomplish a more detailed detection capability. The experimental results show that KBN-YOLOv5 demonstrates better performance with other mainstream algorithms in terms of combined performance of detection accuracy and model computation. Compared to baseline model (YOLOv5), the Recall and mAP values of KBN-YOLOv5 reach 92.3% and 96.0%, which improve 6.7% and 5.4%, respectively. In addition, KBN-YOLOv5 has good robustness, and the model computation is reduced by 6.7% while detection accuracy of the model is improved.

Path planning is an important technique for mobile robots, which can effectively map out a safe path, whether in an open field or in a complex and changing environment. To address the problems of many inflection points and irregular routes in traditional path planning algorithms for mobile robot, this paper proposes a path planning approach that incorporates the improved A* algorithm and the dynamic window approach. First, the algorithm is based on the jump point search algorithm, which extends path finding by finding jump points, thus reducing unnecessary nodes and ultimately achieving the effect of shortening path finding time. Then, the generated paths are smoothed using the gradient descent method to make them smoother and more natural; finally, the improved A* algorithm and the dynamic window approach are fused. This makes the planned path smoother and the obstacle avoidance more accurate, which not only reflects the good path planning ability of the mobile robot, but also meets the practical needs of the mobile robot.

The development of lace has a profound impact on the economic development and new employment in some regions, so the demand for lace is increasing year by year. However, the quality inspection problem has been troubling the whole lace industry. In order to solve this problem, this paper uses the fingerprint identification method to extract and match the details of feature points to detect the defects of lace images. Firstly, a series of preprocessing is carried out between the template and the image to be matched, including normalization, distinguishing the initial foreground and background color, removing the void and burr, image refinement, etc. Secondly, two kinds of detail feature points, namely, endpoints and bifurcation points, are extracted from the thinning graph and the feature information of lace detail points is counted. Then, three kinds of pseudo feature points, short line, outlier and ring, are removed according to the process of distance and ring method. Finally, the Euclidean distance is used to calculate the distance between each step length of the same feature point in the two images and the current feature point, and the distance is saved. If the coordinates of the pixel points stored within the synchronous length are basically the same, it means that the two images are the same. The experiments show that the proposed algorithm is superior to the compared algorithms in terms of both recognition rate and algorithm speed, which meets the requirements of small and medium-sized lace enterprises.