Object Detection With Class Imbalance Based On Knowledge Distillation And Data Augmentation Methods

Object detection is one of the most basic research problems in the field of computer vision,which is used to detect instances of a certain kind of semantic objects(such as people,flowers or buildings,etc.)in digital images and videos.It has a wide range of application scenarios,including intelligent monitoring,urban planning,precision agriculture,automatic driving,unmanned aerial vehicle scene analysis,etc.Object detection tasks can be divided into object classification and location.The object classification determines whether the original image contains the object category of interest,and outputs the probability label with scores.The object location determines the position and size of the target and outputs its square bounding box.At present,many object detection methods based on manual features and deep neural network features have been proposed.However,it is noted that these widely used object detection methods do not consider the problem of object class imbalance(the number of samples of some object classes is much larger than that of other object classes),which will seriously affect the detection effect.Therefore,this paper proposes a object detection method with class imbalance based on knowledge distillation,and the specific research contents are as follows:Firstly,in view of unbalanced object categories,knowledge distillation strategy is used to change the number of samples of each object category in the original data set,so that the number of samples of each category in the newly generated data set tends to be balanced.Then,based on the data augmentation algorithm,according to the sample number of each category determined by the above knowledge distillation strategy,samples are collected from the original data set to generate a new sample data set,that is,excess samples are removed by random sampling for a large number of object categories.The traditional data augmentation methods such as rotation,noise,mirroring and generating adversarial network are used to generate new samples for a small number of object categories.Finally,benchmark object detection algorithms such as YOLOv4 and YOLOv5 are used to classify and locate objects on the new balanced sample data set.Further,in order to verify the effectiveness of the proposed method combining knowledge distillation and data augmentation strategy in the problem of categorically unbalanced object detection,an unbalanced sample flower object detection data set was collected in this paper.The data set contained 20 different types of flowers,and the number of samples of all kinds of flowers varied greatly.The tulip contained 701 image samples at the most.And the least iris only had 34 images.The experimental results on this data set show that the average accuracy of performance indicators of YOLOv4 and YOLOv5 object detection models improved by 2.3% and 2.1%,respectively,compared with the original data set by using the proposed knowledge distillation and traditional data augmentation methods.By using the proposed knowledge distillation and generative adversarial network data augmentation methods,the average accuracy of the performance indicators of YOLOv4 and YOLOv5 object detection models are improved by 2.9% and2.5%,compared with the original data sets.In addition,we also published the PASCAL VOC category imbalance data set(It contains 20 classes,in which the ratio of the number of categories of "people" and the number of categories of "sheep" is 2008: 96).The verification experiment of the universality of the proposed method was carried out.The experimental results showed that the average accuracy of the performance indicators of the YOLOv4 model were improved by 2.6% and 3.1% compared with the original data set,and the average accuracy of the performance indicators of the YOLOv5 model was improved by 2.7% and 3.3% compared with the original data set.Therefore,combining knowledge distillation strategy with data augmentation method to deal with class imbalance in object detection can effectively improve the detection accuracy of the model.