The Research On Key Technologies Of Vision Inspection For Automatic Compressor-grabbing

In recent decades,China has taken the lead on the road of green development.As a country with a large population,China has a large demand for electrical equipment,resulting in a numerous amount of e-waste that contains rich renewable resources.At present,the ability of recycling waste refrigeration equipment in the recycling industry is relatively low,for the core component "compressor " in it has a wide variety of complex shapes,and the manual handling of it in the present recycling process is extremely inefficient.Therefore,how to improve the efficiency of transferring compressors has become an urgent problem to be solved in the recycling industry.At present,several solutions based on computer-vision have been applied to the automatic grabbing of compressors,but none of them have a good performance.In order to design a visual inspection scheme for automatic compressor grabbing that meets actual industrial needs,this thesis investigates and analyzes the actual needs of the recycling industry and the basic principles of designing a visual inspection scheme,dividing the whole problem into the target detection and attitude estimation of the compressor.Firstly,the thesis illustrates the limitations of traditional image processing methods in target detection through relative experiments.Based on the sparsify-training method,the initial YOLOv3 network is channel-level pruned and slimmed for industrial application scenarios,thus the size of the model is compressed by 43.87%,the inference time is shorten by 26.12% and the FLOPs is reduced by 42.30%,while m AP only reduces0.007.On this basis,in order to improve the image quality of the compressor region,a highlight removal method based on a three-channel value mask is proposed,which provides a good image basis for the subsequent attitude estimation module.Secondly,this thesis proposes a foreground-segmentation method based on probability density function estimation,which extracts the foreground part contains the compressor relatively completely in the image.Next,using principal component analysis and the line detection method based on Hough transform,the foreground part is fixed and completed.With data mining and analysis of the foreground part,the main measurements of the compressor are figured out and obstacles attached to the surface of the compressor are detected,which means a series of key point information is obtained.Next,after the introduction of the four coordinate systems of the monocular camera,this thesis has completed the calibration and coordinate conversion work.So far,this thesis has realized of the key technology of visual inspection for automatic grabbing of compressors.Finally,based on the above content,this thesis deploys an actual experimental environment,and conducts visual inspection experiments on 78 compressors which are chosen at random;the detection rate of obstacles on the surface of compressors reaches94.45%,and the average detection time of a single compressor is 1.28 seconds.Obviously,both the accuracy and the efficiency are high,which verify the feasibility of the scheme in this thesis.In summary,the thesis proposes a visual inspection scheme for automatic grabbing of compressors,and the high accuracy and efficiency of the proposed scheme are verified by experiments.