ハイブリッド型転移学習による物体検出における学習の効率化

摘要

In the detection of human from image using statistical learning methods, the labor cost of collecting training samples and the time cost for retraining to match the target scene are major issues. One method to reduce the work involved in sample collection is transfer learning based on boosting. However, if there is a large change between the auxiliary scene and target scene, it is difficult to apply the transfer learning.We therefore propose a hybrid transfer learning method in which two feature spaces are prepared, one of feature obtained by transfer and another of full feature search that is the same as retraining. The feature space is selectively switched on the basis of the defined training efficiency. The proposed method improving accuracy up to 8.35% compared to conventional transfer learning while accelerating training time by 3.2 times faster compared to retraining.%統計的学習法に基づく人検出では，学習サンプル収集に伴う人的コストと特定シーンに合わせた再学習のための時間的コストが大きな問題である.サンプル収集を省力化する手法としてBoostingに基づく転移学習が提案されているが，事前学習シーンと特定シーンの間に大きな変化がある場合，その変化に適応するのは困難である.そこで本研究では，転移によって得られる特徴量と，再学習と同様の全特徴探索の2つの特徴空間を用意し，定義する学習効率に基づいて特徴空間を選択的に切り替えるハイブリッド型転移学習を提案する.提案手法は転移学習の持つ少数サンプルでの高速な識別器構築を特長として持ち，再学習に比べ3.2倍以上高速且つ，従来の転移学習に比べ最大で8.35%の高精度化を実現した.