Research And Implementation Of Target Recognition And Counting Technology Based On Computer Vision

Along with the depth of industrial informatization,the application of computer vision in various industries continues to expand,and as an important part of the industrial chain,the demand for intelligent information acquisition continues to expand.Traditional collection work is done manually,which is time-consuming and inefficient,and has become a painful problem affecting the high-quality operation of the whole industry.In recent years,deep learning algorithms have injected new power into various technologies,and the deployment of algorithm services through servers can build the cornerstone of digital information collection,which is of practical significance to the development of industrial informatization.Information in the industrial chain comes from many aspects,label information and quantity information are two of them.In this thesis,we study text label recognition and target counting algorithms,and applies them to information collection of intelligent devices and factory materials.The recognition of text labels in complex backgrounds has problems such as similar color and light changes,and the existing scene text detection methods are ineffective.Due to the frequent changes of construction sites,it is not practical to arrange 3D reconstruction equipment and other methods for target counting,and the statistical objects are stacked and placed with occlusion,and the existing computer visionbased counting methods gather on flat targets and lack research on stacked target counting.To improve the label recognition effect,in this thesis,we propose an optimized network model based on DB-net,adding single character detection to strengthen the text display,thus increasing the probability of positively determining text and compensating for the defects of missed detection caused by background complexity.The corresponding postprocessing is also designed for information collection objects.Experimental results show that the detection accuracy of the optimized network is significantly improved,and it also plays an optimization effect on the subsequent text processing.To solve the problem of stacked object counting,in this thesis,we propose an algorithm for target counting in monocular images,using the ideas of distribution matching and optimal transmission,adding attention mechanisms to focus on specific features that assist in counting,optimizing the planar density map,and realizing the expansion of planar to three-dimensional space through a depth prediction network,and finally integrating to complete the quantity prediction.Through experimental verification,the algorithm predicts the results with small errors and can meet the application requirements of real industrial environments.In this thesis,we design and implement the embedded application of text label recognition algorithm and target counting algorithm in the information digital collection system to provide intelligent device information collection and factory material information collection functions,including requirement analysis,architecture design,function implementation,and test analysis.The application of the information digital collection system in the industrial chain helps to release manpower,improve efficiency,reduce risks,and provide strong support for the intelligent development of the industry.