基於YOLO深度學習用於小型漂浮物檢測的新型卷積演算法
No Thumbnail Available
Date
2023
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
海洋中的不當廢棄物已導致全球危機，為了緩解這個問題，要在海洋及河流的廢棄物到達環境負荷上限之前對其進行檢測和清理，本研究提出了一種基於 YOLOv4 的算法來檢測河流中的漂流廢棄物，算法結合了改進後的RegP池化層並添加到空間金字塔中的池化層與減少輸出部分的檢測層，以改進特徵提取並防止丟失重要或微小細節，並且針對微小的物品進行檢測。實驗結果中評估了本研究的方法在 FloW和Pascal VOC資料集上的性能，與現今的最先進的技術相比，結果表明提出的方法具有更好的mAP準確率，具體來說，在FloW上分別提升了7.91%和11.36%，並且也與多個在漂流廢棄物檢測的先進方法進行對比，獲得了最佳的準確率，在Pascal VOC上的實驗證實了本研究的方法在不同尺寸大小的物件上的有效性，最後測試了在WIDER FACE上對小尺寸的人臉進行檢測實驗，在準確率上也有一定的提升。本研究提供了一個有前途的解決方案，有助於檢測和清除河流中的廢棄物。
Improper marine debris in the oceans has led to a global crisis. To address this issue, it is essential to detect and clean up waste in the oceans and rivers before reaching the environmental threshold. In this research, we propose a YOLOv4-based algorithm for detecting drifting marine debris in rivers. The algorithm integrates an improved RegP pooling layer added to the spatial pyramid pooling layer and reduces the detection layer's output to enhance feature extraction and prevent the loss of crucial or small details. The algorithm also focuses on detecting smaller objects. We evaluate the performance of our proposed method on the FloW and Pascal VOC datasets. Compared to state-of-the-art techniques, the results demonstrate that our approach achieves higher mAP accuracy, with an improvement of 7.91% and 11.36% on FloW. Additionally, it outperforms several advanced methods for marine debris detection. Experimental resultson the Pascal VOC dataset validate the effectiveness of our approach for objects of varying sizes. Lastly, we conduct experiments on the WIDER FACE dataset to detect small-sized faces, which also show promising improvements in accuracy. This study offers a promising solution for detecting and removing waste in rivers, contributing to addressing the global marine debris crisis.
Improper marine debris in the oceans has led to a global crisis. To address this issue, it is essential to detect and clean up waste in the oceans and rivers before reaching the environmental threshold. In this research, we propose a YOLOv4-based algorithm for detecting drifting marine debris in rivers. The algorithm integrates an improved RegP pooling layer added to the spatial pyramid pooling layer and reduces the detection layer's output to enhance feature extraction and prevent the loss of crucial or small details. The algorithm also focuses on detecting smaller objects. We evaluate the performance of our proposed method on the FloW and Pascal VOC datasets. Compared to state-of-the-art techniques, the results demonstrate that our approach achieves higher mAP accuracy, with an improvement of 7.91% and 11.36% on FloW. Additionally, it outperforms several advanced methods for marine debris detection. Experimental resultson the Pascal VOC dataset validate the effectiveness of our approach for objects of varying sizes. Lastly, we conduct experiments on the WIDER FACE dataset to detect small-sized faces, which also show promising improvements in accuracy. This study offers a promising solution for detecting and removing waste in rivers, contributing to addressing the global marine debris crisis.
Description
Keywords
物件檢測, 深度學習, YOLO, 空間金字塔池化, Object Detection, Deep learning, YOLO, Spatial Pyramid Pooling