Small Target Detection Methods Based On Deep Learning

Object detection is the basis of many computer vision applications,such as instance segmentation,human keypoint extraction,face recognition,etc.It combines the two tasks of object classification and localization.At present,the research hotspots of object detection are mainly focused on methods based on deep learning,while there is still a lot of research space to deal with the imbalance of positive and negative samples and multi-scale detection problems.Therefore,Solutions to these two types of typical problems in object detection are proposed in this paper,which are proved to be effective by experiments.The main contents of this work are as follows:?1?Research and experiments on imbalance of positive and negative samples in object detection.The imbalance of positive and negative samples in object detection is caused by the training data set.The number of targets contained in any image in the training data set is generally small,while the number of negative samples containing background is often rich,which brings about the problem that the ability of object detection framework to detect negative samples is stronger than that of positive samples.Focal Loss solves this problem by improving the classification loss function,but Focal Loss brings additional hyperparameters.Although it solves the problem of imbalance between positive and negative samples,it also brings the problem of adjusting hyperparameters.Based on the idea of Focal Loss,a novel classification loss function SCE is proposed in this paper,which is similar to Focal Loss but contains no super parameters.The final experiment shows that SCE achieves the results close to Focal Loss,and the AP difference between SCE and Focal Loss is basically within 1%,although it fails to surpass Focal Loss's performance,however,SCE omits the problem of adjusting the super parameters.?2?Research and experiments on multi-scale detection in object detection algorithms.The multi-scale detection problem in object detection mainly manifests as a poor detection performance on small targets.FPN is a powerful means to deal with multi-scale detection problems.In this paper,a new Dense FPN structure is designed based on FPN.Dense FPN removes the upsampling process in FPN,and adds successively decreasing convolutional layers after each feature extraction layer.The experiment shows that Dense FPN has a good performance in dealing with multi-scale detection problems.The values of the three evaluation indicators of AP_S,AP_M,and AP_L all exceed the original FPN by about 1%,which shows the excellent performance of Dense FPN in dealing with multi-scale detection problems.?3?The research of small target insensitivity is conducted from the perspective of training data set.Although the imbalance of positive and negative samples and the fusion of multi-scale features are some of the factors that affect the performance of small target detection in object detection,these factors all have an indirect effect.However,the small number of small targets in training data set and the lack of semantic information are the direct reasons for the poor detection performance of small targets in object detection.In this paper,a specific small target data set is designed based on the original MS COCO data set,the image samples in the data set only contain small targets,and do not include targets of other sizes.The experiments in this work show that the small target data set will only have a significant performance improvement on the AP_S in the object detection evaluation index,while the values of other evaluation indexes will produce some losses.At last,we mix the small target data set with the original MS COCO data set,and then use the mixed data set to train all kinds of obejct detection framework.Experiments show that using the mixed data set can improve the values of AP,AP₅₀ and AP₇₅,as well as the values of AP_S,AP_M and AP_L.At the end of the thesis,we conducted a fusion experiment study on the designed loss function,multi-scale feature extraction structure and small target data set.The effectiveness of the SCE loss function,Dense FPN,and small target data set are proved by the results in the six evaluation indicators of AP,AP₅₀,AP₇₅,AP_S,AP_M and AP_L.And then a more robust and efficient small target detection architecture is designed,which improves the small target sensitivity of common target detection.