







摘要:針對Yolov3-Tiny算法在加油站監控場景檢測時由于數據特征提取不充分而導致檢測精度低、 漏檢率高等問題, 提出一種基于加油站場景的Misp-YOLO(You Only Look Once)目標檢測算法。首先引入Mosaic數據增強算法, 使圖片包含更多特征信息; 其次使用InceptionV2和PSConv(Poly-Scale Convolution)多尺度特征提取方法提升網絡多尺度預測能力; 最后結合scSE(Concurrent Spatial and Channel ‘Squeeze amp; Excitation’)注意力機制, 重構主干網絡輸出特征。實驗結果證明該算法具有較高檢測準確度, 并且檢測速度滿足實際需求。優化后的算法性能得到極大提升, 可推廣應用于其他目標檢測中。
關鍵詞:目標檢測; YOLO算法; 特征提取; 注意力機制; 多尺度預測
中圖分類號: TP183 文獻標志碼: A
Misp-YOLO: Gas Station Scene Target Detection
LIU Yuanhong, CHENG Minghao
(School of Electrical Information and Engineering, Northeast Petroleum University, Daqing 163318, China)
Abstract:In order to solve the problem that Yolov3-Tiny algorithm has insufficient feature extraction in gas station monitoring scene detection, which results in low detection accuracy, a new target detection algorithm based on gas station scene is proposed. This method first introduces Mosaic data enhancement algorithm to make the picture contain more feature information. Secondly, InceptionV2 and PSConv(Poly-Scale Convolution) multiscale feature extraction methods are used to improve the network multiscale prediction ability. Finally, combined with the scSE(Concurrent Spatial and Channel ‘Squeeze amp; Excitation’)attention mechanism, the output characteristics of the backbone network are reconstructed. The experimental results show that the algorithm has high detection accuracy and the detection speed meets the actual needs. The performance of the optimized algorithm is greatly improved and can it be applied to other target detection.
Key words:target detection; you only look once (YOLO) algorithm; feature extraction; attention mechanism; multi-scale prediction
0 引 言
加油站操作規范日常監測可極大地降低發生事故的風險, 是確保加油站安全穩定運營的重要環節。對加油站操作規范檢測, 目前已有許多相關研究報道[1-3]。傳統檢測方法首先根據數據特點(如形狀、 大小和紋理等), 通過人工確定需要提取的重要特征[4-5], 設計出對應的特征提取模型, 最后將提取的特征輸入到適合的分類器中[6-7], 實現對目標檢測。但這類方法主要通過人工經驗選擇特征, 無法自動從數據中學習到重要特征, 提取的特征不夠完備, 影響后續的識別精度。
深度學習算法屬于一種監督學習方法, 能根據實際與期望輸出的差異主動調整網絡參數, 實現顯著特征的自動提取, 而且采用基于深度學習的目標檢測精度通常高于傳統方法的檢測精度, 因此深度學習方法較傳統檢測方法更具工程應用潛質。……