Industrial Equipment Object Detection Based On Faster R-CNN In Complex Scenes

Industrial equipment detecting is a very important part of industrial production.It not only requires a lot of manpower and material resources,but also has hidden dangers to personnel safety.At the same time,some industrial equipment cannot be tested normally.However,traditional manual feature-based object detection takes a long time and has poor robustness,and cannot quickly and accurately detect industrial equipment,especially when dealing with multiple objects,the problem is more prominent.With the rapid development of computer vision,convolutional neural networks have been deeply studied and gradually applied to equipment detection in industrial engineering.The object detection of industrial equipment based on convolutional neural network can not only replace manual on-site inspection,but also is safe and reliable.It can also improve the accuracy and speed of detection,accumulate data materials,and conduct initial exploration for the realization of an unmanned chemical factory.Faster R-CNN is a classic algorithm for object detection in recent years,and it has a high accuracy rate for conventional target detection.However,in actual industrial engineering,due to specific location requirements,different types of industrial equipment will have multiple scales,overlap of occlusions,and rotation tilt,which affects the detection accuracy of industrial equipment.Therefore,this article improves the Faster R-CNN algorithm for industrial equipment detection in complex scenarios.Aiming at the detection of industrial equipment in multi-scale and occlusion overlapping scenes,this paper first selects the lightweight Mobile Net V2 network with deep convolution and separable.The overall effect is similar to standard convolution,but it can greatly reduce the calculation parameters,reduce memory usage and time overhead,and ensure detection accuracy.Secondly,build the Mobile Net V2 feature pyramid network on this basis.The feature map with lower resolution and stronger semantic information is combined with the feature map with higher resolution and weaker semantic information,so that each layer contains strong semantic information,and each layer is predicted separately.Finally,a multi-stage penalty Non-Maximum Suppression algorithm is proposed,which attenuates the confidence of the candidate frame through the introduced multi-stage penalty factor,and improves the missed detection caused by the direct deletion of the candidate frame.Experiments on industrial equipment detection have shown that the improved algorithm improves the accuracy of industrial equipment detection in multi-scale and occluded overlapping scenes.Aiming at the detection of multi-category industrial equipment in the rotating and tilting scene,this paper first proposes a specific rotating anchor scheme,which clusters the training set of multi-category industrial equipment through the k-means clustering algorithm to obtain a specific aspect ratio information,combined with different rotation and tilt angles,generates an anchor frame with multiple aspect ratios and multiple rotation and tilt angles.Secondly,the overlapped part of the candidate frame is divided into triangles using the rotating Intersection over Union algorithm.Calculate the area of each part of the triangle separately and then add them to get a precise intersection ratio.Finally,through the multi-scale Region of Interest pooling layer,the feature maps of different sizes are normalized,and the judgment conditions of the positive and negative anchors and the multi-task loss function are modified at the same time.The detection experiments on multi-category industrial equipment have shown that the improved Faster R-CNN based industrial equipment object detection algorithm has a better detection effect than other algorithms.It can successfully detect multi-category industrial equipment in complex scenes such as multi-scale,occlusion overlap,and rotation and tilt,which effectively improves the missed and false detection situations,and improves the accuracy and speed of industrial equipment detection.