天天看點

21-ICME-FINE-GRAINED IMAGE RETRIEVAL VIA MULTIPLE PART-LEVEL FEATURE ENSEMBLE

MPFE(MULTIPLE PART-LEVEL FEATURE ENSEMBLE)

21-ICME-FINE-GRAINED IMAGE RETRIEVAL VIA MULTIPLE PART-LEVEL FEATURE ENSEMBLE

AAPD基于注意激活的零件檢測器

特征圖:

21-ICME-FINE-GRAINED IMAGE RETRIEVAL VIA MULTIPLE PART-LEVEL FEATURE ENSEMBLE
21-ICME-FINE-GRAINED IMAGE RETRIEVAL VIA MULTIPLE PART-LEVEL FEATURE ENSEMBLE

1 CAD(通道式注意力檢測):

21-ICME-FINE-GRAINED IMAGE RETRIEVAL VIA MULTIPLE PART-LEVEL FEATURE ENSEMBLE

MP最大池化k-max-pooling

MLP 兩層全連接配接層(編碼通道間關聯資訊)【μ是c維向量?經過兩層全連接配接還是c維?】

softmax标準化

激活圖 :

21-ICME-FINE-GRAINED IMAGE RETRIEVAL VIA MULTIPLE PART-LEVEL FEATURE ENSEMBLE

【原理類似MA-CNN,相當于聚類;形式類似SE-net,給特征圖賦權重】

2 PS(部件選擇):

SCDA

21-ICME-FINE-GRAINED IMAGE RETRIEVAL VIA MULTIPLE PART-LEVEL FEATURE ENSEMBLE
21-ICME-FINE-GRAINED IMAGE RETRIEVAL VIA MULTIPLE PART-LEVEL FEATURE ENSEMBLE

MPFE多零件級特征內建

21-ICME-FINE-GRAINED IMAGE RETRIEVAL VIA MULTIPLE PART-LEVEL FEATURE ENSEMBLE

Loss

21-ICME-FINE-GRAINED IMAGE RETRIEVAL VIA MULTIPLE PART-LEVEL FEATURE ENSEMBLE

backbone:Inception network 有batch normalization 預訓練,沒有logits layer

是不是輕量級網絡更适合提取了局部特征的檢索模型,防止過拟合?

為了高效獲得局部區域,輸入圖像大小為 512 × 512,然後裁剪到454 × 454.

繼續閱讀