Target Detection Method Based On Improved YOLO Algorithm

Object detection,as an important branch in the field of computer vision,has a wide range of applications in many occasions,such as information collection,military fields,and intelligent human-computer interaction.However,because the objects in real life are easily affected by many factors such as light,occlusion,and environmental changes,object detection algorithms based on traditional manual design features have been unable to obtain good detection results.In recent years,as a powerful tool,deep learning has taken advantage of its self-learned features under big data,which makes people research in the field of object detection further.Compared with traditional object detection algorithms that rely on manual calibration features,object detection algorithms based on deep learning have obvious advantages in terms of quantity and performance.However,due to the complexity and uncertainty of the real scene,it adds some difficulties to the actual detection process.Therefore,object detection algorithms based on deep learning still have great challenges.Based on the combination of the traditional object detection algorithm theory and the application of object detection in the field of deep learning,this paper studies the object detection algorithm based on deep learning,and the main research contents include the following aspects:1)For the original YOLOv3 detection model,the detection accuracy is not high,especially on the problem of poor detection of small objects.To improve detection results,the network structure of Pyramid Net + SE is proposed as the backbone of the model.First,the SE module is used to implement the attention mechanism on the channel,so that the model can fully extract features from the channel dimension.Then,the AE layer is used to solve the problem of positioning errors;2)Aiming at the problem of serious imbalance between the positive and negative samples of the YOLOv3 detection model,based on the original loss function,a GHM loss function was introduced to reduce the weight of simple negative samples in the training process,thereby improving the detection accuracy;3)In order to avoid being eliminated by the non-maximum suppression algorithm due to the detected objects that overlap too much,use Soft-NMS instead of the NMS algorithm;4)In order to improve the robustness and generalization ability of the improved YOLOv3 algorithm,data enhancement work has been performed on the data in various ways(such as changing color channels,geometric distortions,and monomorphic transformations).