Small Target Modified Car Parts Detection Based on Improved Faster-RCNN

Because the small target image has fewer pixels, it is prone to miss detection and error detection; the target detection algorithm needs to be improved and enhanced to deal with the problem about the small target detection in specific scenarios. In this paper, small target detection algorithm is applied to the new field of illegally modified vehicle detection to reduce the workload of traffic management. The commonest modification is to install a rear wing and change the contact angle between the hub and the ground. This paper proposes a method to detect modified car parts based on improving Faster-Rcnn, in order to detect two modifications described above. On the basis of the original Faster-Rcnn, using multi-scale training and increasing the number of Anchors to enhance the robustness of the network in detecting targets of different sizes, and introducing the Soft-NMS algorithm to replace the NMS algorithm, to solve the problem of partial overlap between two targets when the distance is close, the possibility of missed detection of targets with low confidence and bounding boxes with high confidence scores are not always more reliable than bounding boxes with low confidence. Experiments show that compared with the original Faster-Rcnn, the detection accuracy is increased by 4.6%, and the model has a certain generalization ability and robustness.