Research On Object Detection Algorithm Based On Deep Learning

With the development of deep learning,computer vision is widely used.As its basic algorithm,object detection has become a research hotspot.Therefore,it is of great significance to study the target detection algorithm to get higher accuracy.In this paper,based on public data sets,we first explore the influence of different settings of box regression loss function on the positioning accuracy of object detection.Then we try to add the attention mechanism into the non maximum suppression algorithm of the filter accuracy box.Finally,we focus on the problem of mismatching in the corner object detection algorithm and propose an improvement to delete the big box caused by mismatching.The main innovations are as follows:1.Two methods are proposed: the redefined generalized intersection union ratio loss function and the improved distance based loss function.The first method focuses on the intersection and coincidence between the prediction box and the real box,while the second method focuses on the distance change after the prediction box and the real box move in the regression process.The two methods are introduced into the classical algorithms of one-stage and two-stage object detection,and the accuracy is compared with the original loss function.The results show that the two improved methods can improve the box regression accuracy which shows that the algorithm has generalization ability.2.An attention mechanism non-maximum-suppression algorithm is proposed.The attention mechanism which is widely used in natural language processing and image processing is introduced into the non-maximum-suppression algorithm.Only the score sorting method is used as the standard to remove the redundant box.The score and location information are combined with the attention algorithm.The experimental results show that the proposed algorithm can improve the detection accuracy.3.The model of eliminating inclusion box is proposed.To solve the problem of corner matching error in corner detection network,a model is proposed to improve it.The box with score more than 0.5 is checked and classified,and the large box with error matching is removed by using the model.Experiments on public data sets show that the algorithm achieves good visual effect and improves the accuracy of corner detection algorithm.