Research On Single-stage Object Detection Algorithm Based On Convolutional Neural Network

In the field of computer vision,deep learning based object detection algorithms are one of the core topics.Object detection identifies and localizes objects in a picture,and it is the basis for many classification tasks.Early object detection models generally recognize objects in two stages,which requires a large amount of computational resources.The single-stage object detection models proposed afterwards run faster but often have a significant degradation in recognition accuracy and are difficult to apply.With the development of object detection models in recent years,the recognition accuracy of single-stage object detection models has been greatly improved,while they are used in several fields due to their faster running characteristics.Therefore,the study of single-stage object detection models is necessary.The work of this thesis investigate and improve the single-target detection model for several problems such as unsynchronized classification prediction value and prediction bounding box quality,training underfitting,and mismatch between classification prediction value and prediction bounding box quality in non-maximum suppression(NMS).The main research is as follows:(1)In general,the models bounding box regression is trained independently from the classification prediction,and the level of classification confidence is not necessarily related to the quality of the prediction bounding box.In order to improve the connection between classification confidence and predicted bounding box,a weight coefficient is added to the loss function of the bounding box regression in the training.This weighting factor assigns a different weight to each prediction bounding box according to the classification confidence level,and the weighting is higher when the classification confidence level is higher.Similarly,Io U weights are added to the classification loss function to improve the training effect of the classification confidence corresponding to higher quality prediction bounding boxes.The use of weight coefficients by the model increases the association between classification confidence and prediction bounding box quality.The models were retrained using this weighting factor on several models such as PAA,FCOS,and ATSS and validated on the MSCOCO dataset.All of these models showed an improvement in m AP of about0.4%,demonstrating the effectiveness of the method.(2)The Weighted-NMS method is optimized.The general NMS algorithm directly removes the parts of the prediction bounding boxes with high overlap,while the Weighted-NMS method calculates a weighted average of these overlapping prediction bounding boxes.Firstly,the Weighted-NMS method is used on the model,but it is found to be ineffective or even negative.The Weighted-NMS method is improved after the study.When Weighted-NMS calculates the weighted average,not all the overlapping prediction bounding boxes are used directly.Instead,the top prediction bounding boxes with the largest overlap are selected according to their degree of overlap,and only the weighted average is calculated for them.It was verified that there was an improvement of about 0.3% in m AP for different models.(3)Optimize the way of the learning rate drop during training.The general SGD decreases the learning rate directly by a factor of n,which leads to an undertrained model and is more prone to overfitting in the same situation.The improved SGD learning rate method does not drop directly by a factor of 10,but in a stepwise way.This improvement makes the model more accurate in detection at the same number of iterations.The modified SGD was experimented on several models,and there was an improvement of about 0.3% in m AP on all of them.After several experimental verifications,our improved method is easy to implement on different models.The implementation is also simple,with only a few changes to improve the detection accuracy of the model.