| Object detection is an essential research direction in the field of computer vision.It is defined as that the computer automatically recognizes the category of a specific object in a given image or video and its specific position in the image or video.Object detection is an important theoretical basis for high-level vision applications such as segmentation,scene understanding,object tracking,image annotation and event detection.It is also widely used in robot vision,video surveillance,autonomous driving,new retail and other application scenarios.Small objects are those with a lower resolution or a smaller size than a picture or video,that is,a small target has limited features and is more susceptible to interference from background noise during feature extraction.In addition,detecting small objects is more complicated,so it cannot detect objects very fast.At present,deep learning-based object detection methods are in an absolute leading position in this field,so researching small object detection algorithms has not only important academic value,but also practical application value.This thesis improves the shortcomings of the object detection algorithm SSD in small object detection.The main research work is as follows:1.A fusion method based on dilated and transposed convolutions to enhance small object features is proposed.The algorithm uses a dilated convolution module on shallow features to enhance the receptive field while maintaining its resolution to fit the size of the default boxes;transposed convolution is used on deep features to enlarge feature resolution to ensure its combination with shallow features to introduce high-level semantic information into shallow layers for improving the effect of small object detection.The proposed algorithm can detect small objects effectively.2.A small target detection method based on object detection feature extraction network DetNet is proposed.This algorithm replaces the backbone of the FSSD algorithm with DetNet,and modifies its feature fusion layer to better match the use of the basic network and improve the ability to detect small objects.Experimental results on PASCAL VOC and MS COCO datasets show that the algorithm can detect small objects quickly and efficiently.3.A fast and efficient method for small object detection based on model pruning is proposed.The algorithm first modified the feature fusion layer of FSSD to increase the number of channels for small target detection layers.At the same time,it added sparse training by adding a batch normalization layer.After training,pruning was performed according to the channel scaling factor.Then fine tuning was continued to obtain the final result,pruned model.The experimental results show that the modified model and the pruned model have achieved a good balance in speed and accuracy,and the pruned model can achieve real-time detection. |
