Object Detection Algorithm Based On Convolutional Neural Network And Its Application

In recent years,deep learning has achieved a series of remarkable research results in the fields of computer vision,speech recognition,natural language processing and medical image processing.In different types of deep neural networks,convolutional neural networks(CNN)have been widely studied for their powerful image feature extraction capabilities,not only reflected in the prosperity of academic research,but also has a huge practical impact and commercial value for related industries.With the rapid growth of labeled sample data sets and the significant increase in graphics processing unit(GPU)performance,CNN has achieved outstanding results in various tasks in the field of computer vision.As the basis of image understanding,Object Detection has always been an important research direction in the field of computer vision.It attempts to accurately locate the location area of the object object in the image and determine its category.It is widely used in daily life scenes such as intelligent video surveillance,virtual reality,human-computer interaction and automatic driving.In recent years,although convolutional neural network-based object detection algorithm has developed rapidly,it still faces many difficulties and challenges in practical application,such as too many weight parameters of model,long time of object detection,difficult to balance speed and accuracy,difficult to detect small objects,changeable natural scenes and so on.In view of the above problems,this paper improves the Faster R-CNN object detection algorithm based on candidate region ideas and the YOLO v3 object detection algorithm based on regression theory to meet the requirements of speed,memory and precision in practical applications:(1)Because the traditional CNN network model has many parameters,large memory occupancy and high computing cost,it is difficult to deploy on mobile devices with limited computing power and memory resources.Therefore,we compress the parameter quantity and reduce the computation cost of the regression-based object detection model YOLO v3 to build a lightweight object detection model MobileNet_YOLO_v3 suitable for embedded devices by using the depth separable convolution laye to replace the standard convolutional layer used by the traditional model.The depth separable convolution reduces the parameter quantity and computational cost of the model by solving the standard convolution integration into deep convolution and point-by-point convolution,and extracting and merging the image features in two steps.The comparison test results show that the use of the depth separable convolutional layer to replace the standard convolutional layer greatly reduces the parameter quantity of the model,reduces the memory footprint,and reduces the detection accuracy of the model only within the acceptable range.(2)Aiming at the problem that Faster R-CNN,a object detection model based on the candidate region,has a slow running speed and high computational cost,we propose a smaller and more efficient lightweight two-stage target detection model Light R-CNN.Light R-CNN uses the lightweight network MobileNet v2 instead of the large network Resnet-101 as the feature extraction network,generates feature maps with fewer channels,and the extracted features are then classified and located using an R-CNN subnetwork containing only one fully connected layer.Because the main structure and head structure of Light R-CNN are relatively light,the calculation amount of the model is greatly reduced,the memory required for detection is less,the detection speed is significantly improved,and an effective trade-off between speed and accuracy is achieved.The contrast experiment shows that the newly proposed lightweight object detection model Light R-CNN greatly improves the detection speed of the model and reduces the number of model parameters on the basis of ensuring the detection accuracy.