Specific Object Recognition and Tracking by Cascade Connection of Different Types of CNNs and a Time-series Filter
Takaki Nishio, Takeshi Nishida · 2018
Currently, convolutional neural networks (CNN) are widely used for object recognition, and it is common to use an already trained network as a basis. However, the retraining cost of CNN to recognize a specific object is high in terms of dataset preparation and calculation cost. Thus, we propose instead to add a second CNN specialized in the recognition of the desired target. Further, a dataset pre-filtering method is proposed to train the second CNN for specific object shape recognition. Additionally, a time-series filter is used to stabilize the detection of the target. To confirm the effectiveness of the proposed method, we use recorded videos and evaluate the accuracy of specific carrot shape recognition.