Few-shot object detection model based on transfer learning and convolutional neural network

Abstract

Abstract Object detection is an important research content in computer vision. Faster R-CNN has been widely used in object detection with large samples. Sample scarcity in practice makes it difficult for Faster R-CNN to extract effective feature and obtain high detection accuracy. In this paper, an improved few-shot object detection model is proposed based on Faster R-CNN. Transfer learning algorithm is applied to transfer trained weights of similar source datasets firstly. Then, attention mechanism is introduced and loss function is improved to construct the few-shot object detection model which has powerful generalization ability. Experimental results on PASCAL VOC dataset show that the proposed model can converge faster and the detection accuracy mAP reaches 79.1% which is 14.4% higher than that before the improvement. Compared with SSD and YOLO v3 network, the detection accuracy of the proposed model is also increased 31.4% and 13.1%, respectively.