高性能之輕量級卷積神經網路之設計

Abstract

因深度學習強大的分析能力，其時常被用做影像辨識與物件偵測的工具。時至今日，已有許多基於深度學習方法的著名模型被提出，例如：SENet、EfficientNet、Densenet、MobileNet、ResNet、ShuffleNet、GhostNet、Yolo等。深度學習模型的性能主要可從4個層面進行探討，分別是參數量，資料分析能力、處理資料的速度以及模型的泛用能力。一般而言，模型能在上述4個層面都表現優秀是很困難的。在本論文中，我們設計出一各性能皆優秀的深度學習模型―ExquisiteNetV2。我們選用了15個具公信力的影像辨識資料集以及1個物件偵測資料集進行實驗，並以上述提到的著名模型做為比較對象。我們分別使用兩種不同的權重更新法做實驗，根據實驗結果，無論使用何種權重更新法，在超過一半以上的資料集中，ExquisiteNetV2的分類正確率都是第一名。ExquisiteNetV2的參數量遠少於其他模型，但資料分析能力以及運算速度卻優於其他模型，因此，ExquisiteNetV2是一種高性能之輕量級卷積神經網路，可通用於影像分類與物件偵測之應用。The deep learning is often applied in the object detection and image recognition because of its outstanding ability of analyzing images. So far, many famous deep learning models have been proposed, such as SENet, EfficientNet, DenseNet, MobileNet, ResNet, ShuffleNet, GhostNet, Yolo and so on.There are four types of model performance to be evaluated, namely the amounts of parameters, the ability of analyzing data, computing speed and model generalization. Generally, it is difficult to perform well in all the four types of performance for a model.In this paper, we proposed a outstanding model called ExquisiteNetV2 which is good at all the four types of performance. We test ExquisiteNetV2 and the aforementioned models on fifteen credible datasets. According to the experimental results, ExquisiteNetV2 gets the highest accuracy on over half datasets. Moreover, ExquisiteNetV2 has the fewest parameters and its ability of analyzing data is better than other models. Experimental results show that ExquisiteNetV2 is a high performance lightweight model suitable for image recognition and object detection.