Cyclic Learning of a Frame Downsampler and a Recognion Model in High-Speed Camera Image Recognition

Shigeaki Namiki, Takuya Ogawa, Keiko Yokoyama, Shoji Yachida, Toshinori Hosoi

Abstract: High-speed cameras capture instantaneous changes in dynamic phenomena, enabling new image classification applications such as nonstop visual inspection of dynamically moving objects. However, due to high-speed cameras’ high frame rate, downsampling excess frames or limiting the size of the recognition model is necessary for real-time processing. Previous work has introduced a frame downsampler (a binary classifier optimized to predict in advance whether the output of a recognition model is true or false) and applied the downsampler to retain high-scoring frames for the recognition model. However, further optimization of the recognition model for the sampled data distribution is unexplored and remains sub-optimal. In this study, we propose “cyclic learning” for high-speed camera image recognition. It optimizes the recognition model for the data distribution left by the downsampler and retrains the downsampler based on the updated recognition model. We constructed a dataset of fast-moving objects captured by a high-speed camera to classify object types, and experimental results on this dataset proved that the proposed method outperformed previous studies regarding overall classification performance with the same number of samples and the number of samples required for comparable classification performance.