Research On Object Detection Algorithm Based On Deep Learning

In the context of the era of big data and artificial intelligence,the rapid development of deep learning has led to major developments in computer vision tasks.As a core task in the field of computer vision,object detection has also developed vigorously.But,object detection still faces a large number of challenges,and is always struggling on the road to improve detection accuracy and detection speed.This paper is based on deep learning algorithms.The purpose of the paper is to provide feasible solutions to some challenges in the field of object detection.The main work of this paper is as follows:1.Aiming at the challenge of sample imbalance in the field of object detection,some solutions at this stage are summarized from the perspectives of difficult sample mining,difficult sample generation,and sample loss allocation.An effective solution is proposed from the perspective of anchor boxes and loss.Unreasonable anchor box settings will aggravate the imbalance of positive and negative samples.For this purpose,clustering algorithm is used to cluster the ground truth boxes of the data set to obtain more reasonable anchor box settings.Some negative samples have strong position regression capabilities,even more than the positive samples.These samples are potential positive samples.Directly dividing them into negative samples will further aggravate the imbalance of positive and negative samples.This will affect the positioning accuracy of the detector.For this,the dynamic anchor box learning algorithm is used to divide the sample which combines the regression ability of samples.The loss function is also redefined.Experimental results show that this solution can effectively alleviate the imbalance of positive and negative samples,improve the recall rate of the object,and improve the accuracy of the detector.2.Aiming at the small object detection challenge in the field of object detection,the reasons why small objects are difficult to detect are briefly described.Some solutions of existing research work are summarized from the perspective of scale,context feature information,anchor box configuration,and threshold.Because of SSD's poor detection effect on small objects,a new method is proposed.Small objects have fewer pixels and less available information.For this,a feature enhancement module is proposed,which makes full use of the surrounding information of small objects,and enhances the features of small objects through context information.The number of prediction boxes matched by small objects is small,and the strict threshold setting will make the number of small object's positive samples fewer.For this,the adaptive training sample selection algorithm is adapted.For each real object,the threshold is dynamically selected according to its actual situation to ensure the number of boxes matched to small objects.The experimental results show that the SSD algorithm based on feature enhancement and adaptive training sample selection strategy improves the detection accuracy of small objects significantly.The algorithm has speed and accuracy advantages,and surpasses SSD and related improved algorithms.3.Aiming at the designing attention mechanism challenge in the field of object detection,the principle of attention mechanism and the core idea of main stream soft attention mechanism are introduced.Part of the existing attention mechanism is introduced from the perspective of spatial attention,channel attention and mixing of them.A very lightweight hybrid attention mechanism is proposed,including spatial attention module and channel attention module.The spatial attention module pays more attention to the importance of spatial pixels.For this,it extracts the mean value of all channels at each pixel position,and uses ordinary convolution and dilated convolution to extract the importance of each spatial position,and then generates a spatial position mask.The channel attention module pays more attention to the importance of channel information.In order to reduce the parameters,the channel weight parameters are set by parametric method.We see from the results,the attention mechanism is extremely lightweight and can be easily embedded in the convolutional neural network.It improves the network's attention on important areas and effectively improves the accuracy of the object detection task.Taking the challenges faced by object detection tasks as the starting point,this paper analyzes and summarizes the three major challenges such as sample imbalance,small object detection,and attention mechanism design.The paper proposes related solutions.Experiments verify the effectiveness of these solutions.The relevant research in this paper has certain practical significance for improving the performance of the detector and solving some of the challenges in the field of object detection.