Research On Object Detection Method Based On Convolutional Neural Network

Detection,segmentation and classification tasks are important areas of computer vision research.There is no doubt that object detection belongs to the detection task in computer vision.Although object detection begins to rise,it is a complex process,which needs to face many challenges,and the biggest challenge in object detection is to maintain the balance between speed and precision.In the process of achieving balance,we will encounter various challenges,such as the detected objects have different angles,different sizes,different shapes and are blocked.This series of challenges will affect the precision and speed of detection.Among many region selection strategies,window based sliding is used in traditional object detection methods.The sliding window is lack of pertinence for the selection of object frame,resulting in cumbersome and time-consuming process and weak real-time performance,while the manually designed features have poor robustness and low recognition precision.In recent years,the method based on convolutional neural network(CNN)can effectively extract the low-level and high-level features of images,and its performance in object detection is dazzling.Contemporary object detection methods based on CNN can be roughly divided into one-stage object detection method represented by YOLO series and two-stage object detection method represented by RCNN series.Because "candidate region + feature resampling" is the core operation of the two-stage detection method,the object detection precision of the two-stage method is high and the speed is slow.On the contrary,the precision of the one-stage object detection method is slightly low and the speed is fast.In view of the challenges faced by object detection at this stage,this paper mainly studies from the following three aspects:(1)A object detection model based on multi-level feature fusion is proposed to improve the detection precision of small object objects.The model combines the spatial information of high-level features with the location and detail information of low-level features,so as to strengthen the feature extraction and greatly improve the precision of object detection in small object detection.At the same time,we introduce V module and bidirectional FPN to better promote the flow of information by using bottom-up and topdown dual paths,and enhance the whole feature level by using accurate low-level feature information,so that the model can make more full use of the feature information of multilevel structure.After feature fusion,CBAM attention mechanism is added to make the model pay better attention to the important features of the image,so as to effectively improve the detection precision of the object.(2)A object detection method based on residual block multi-scale fusion is proposed to further improve the object detection precision.This paper proposes a residual block based on dense connection.After each feature fusion,the residual block can be used to supplement the features effectively.Through the introduction of dense connection,we can make better use of high-level and low-level feature information.Among them,rich spatial semantic information mainly exists in high-level features,while rich detail information and location information mainly exist in low-level features.The combination of low-level and high-level complements each other.Thus,the detection precision of objects with different scales can be further improved.(3)A object detection method based on multi-layer receptive field feature jump connection fusion is proposed to reduce the complexity of the network and improve the object detection precision.In order to make the object detection not constrained by a single receptive field,we use the fusion of asymmetric convolution and multi-scale features to build a network with multi-layer receptive fields,and use jump connection to strengthen the reuse of features.By using asymmetric convolution,we can reduce the parameters and complexity of the network model and improve the detection speed.