Research On Object Detection Of Mechanical Parts Based On Deep Learning Of Attention Mechanism

Intelligent manufacturing is one key to the transformation and upgrading of the manufacturing industry.While improving production efficiency and product quality,it also presents new challenges for the object detection system of mechanical parts.Traditional object detection systems of mechanical parts have problems such as large errors of pose detection and low accuracy of category recognition.On the other hand,human visual systems help people easily detect objects in complex environments by virtue of selective attention mechanisms.Inspired by this,this thesis deeply researches the selective attention mechanism and develops an anthropomorphic part object detection system.The main research work includes:(1)A part saliency evaluation algorithm based on robust background detection is applied to detect the part saliency in the visual scene of mechanical parts;the image segmentation method based on Otsu is used to obtain accurate pose information from the part saliency detection results.Experiments show that the proposed algorithm can quickly calculate and determine the pose of object parts in scattered environment.(2)Based on the study of the selective attention model coupled with Hodgkin Huxley neurons and its synchronous dynamic mechanism with the evolution of attention state,the part object selection attention module is designed by introducing position weight of the salient part as the external input of peripheral neurons,the plasticity of neurons and the cumulative time of inhibition between neurons.The designed part object selection attention module improves the rationality of the selection attention algorithm and the detection efficiency of the part.(3)By integrating the saliency detection module,the segmentation module and the selection attention module,a saliency part selection attention system is constructed.The system can select the saliency parts according to the saliency of the parts in the visual scene,and accurately determine their pose.The experimental results show that the system can achieve good selection attention effect on the salient parts in the visual scene of mechanical parts.(4)The recognition algorithm of mechanical parts based on residual network is constructed,and the recognition training and testing are carried out on the dataset of mechanical parts which contains 9 kinds of 2880 images.The test results show that the recognition accuracy of the part recognition algorithm is more than 99% under complex background.(5)Finally,a part recognition module based on residual network is integrated into the salient part selection attention system,which can detect the salient part and its pose,and recognize the part type at the same time.Compared with the traditional detection algorithm,the pose detection error is small and the type recognition accuracy is high,which is closer to the information processing principle of human visual nervous system.