Cascaded Deconvolution-convolution Structure Of Object Detection With Pixels IoU Loss Function

Object detection technology plays a fundamental role in computer vision and is the basis for many advanced vision tasks.At the same time,it is widely used in many scenarios,such as smart cities,smart medical treatment,smart agriculture,etc.With the rapid development of deep learning,object detection make an significant improvement,however,the object detection technology still faces the challenge of more complex and diverse scenarios and object.Aiming at the shortcomings of the existing object detection algorithms,this thesis proposes algorithms with practical significance and originality from network structure and loss function.On the one hand,existing object detection algorithms have always faced the difficulty of detecting small-scale and fuzzy objects.This thesis proposes a cascaded deconvolution-convolution(CDC)structure to enhance features to improve the detection accuracy of objects.CDC expands the resolution of feature maps,meanwhile,keep the crucial information in the feature maps.and to effectively fuse shallow-level and deep-level features and get more rich features,and then adopts a double helix connection(DHC).On the other hand,the bounding box for object detection cannot represent more complex object’s location more perfectly.The existing orientated object detection uses a orientated bounding box to locate more multiangle targets.In the orientated object detection algorithm,there is some gap between the distance loss function’s regression on the orientated bounding box and the intersection of union(IoU)-based evaluation method.In order to eliminate the above gap,this thesis proposes an efficient and simple pixel IoU loss function from the perspective of a single pixel.Finally,in order to verify the effectiveness of our algorithm,we performed a lot of experiments on different mainstream datasets.Experimental results show that our algorithm can significantly improve the object detection performance,especially for small-scale and fuzzy objects.At the same time,our proposed pixels IoU loss function can greatly improve the accuracy of object location,especially for objects with extreme aspect ratios.