| Detecting small object is the largest challenge of object detection.Small object is that the height and width of target are both smaller than 32 pixels in image.Many conditions can influence the small object detection such as illumination changes,color deterioration,motion blur,adverse weather conditions,cluttered background,partial occlusion and so on.Moreover,there are many difficulties for small object detection because of small volume,low resolution and less information of small object.Compare to object with large and middle size,the above situations cause the low performance of the detection networks.To resolve these problem,this thesis leverages the attention mechanism and the method of information fusion to creatively design new structures of detection network and to improve the performance of detection on small object.Attention mechanism can make the detector pay more attention to the areas that the detector want to learn and confine the power on the non-attention areas.The method of information fusion can adequately make use of the mixture of the strong semantic information from deep layers and detailed information from shallow layers.It also can compensate deep layers for the loss of the information of small object with the information from shallow layers.Inspired by these two principles,this thesis conduct the plans as following:Firstly,this thesis designs an attention convolution structure.It takes a convolution kernel and a matrix with constant numbers in feature process to acquire a weight tensor about interesting feature.This weight tensor can promote networks to pour more power to learn the area where network is interested and eliminate the interferential information.As a result,it make the detector for small object performance better.Secondly,this thesis designs a new feature pyramid module.In order to learning appropriate features and the probabilities of the utilization of valid information,it takes advantage of mechanism to weigh the information of origination and reestablishment to refine the way of learning of network.Meanwhile,the structure increase the new path to merge information and fusion method to enhance the expression to improve the performance of the detector of small object.Thirdly,this thesis refine the frame of the feature pyramid module by adding a multiple receptive fields block to scratch more discrimination information between small object and its context.As a result,it extracts more useful information and promote the small detector performance well.This thesis designs attention convolution algorithm 、 double upsample feature pyramid algorithm 、 multiple branches receptive field algorithm.These,three algorithms,perform well on detection.They supply values for detection task and possess practical significance. |
