Small Object Detection in Hyperspectral Images Based on Radial Basis Activation Function

Bofan Wang; Haitao Zhao

doi:10.3788/AOS202141.2311001

Journals >Acta Optica Sinica >Volume 41 >Issue 23 >Page 2311001 > Article

Acta Optica Sinica
Vol. 41, Issue 23, 2311001 (2021)

Small Object Detection in Hyperspectral Images Based on Radial Basis Activation Function

Bofan Wang and Haitao Zhao^*

Author Affiliations

School of Information Science and Engineering, East China University of Science and Technology, Shanghai 200237, China

show less

DOI: 10.3788/AOS202141.2311001 Cite this Article Set citation alerts

Bofan Wang, Haitao Zhao. Small Object Detection in Hyperspectral Images Based on Radial Basis Activation Function[J]. Acta Optica Sinica, 2021, 41(23): 2311001 Copy Citation Text

show less

Fig. 1. Schematic diagram of the radial basis activation function for spectral screening (RBAF-SS)

Download full size

Fig. 2. Attention-based resolution reconstruction network (ABRRN)

Download full size

Fig. 3. Radial basis object output network (RBOON)

Download full size

Fig. 4. Examples of hyperspectral data sets. (a) Small objects; (b) medium objects

Download full size

Fig. 5. Qualitative analysis of spectral selection based on RBAF

Download full size

Fig. 6. Spectral curve of airplane object features

Download full size

Fig. 7. Experimental results of the proposed method are compared with four approaches with sigmoid based object output layers. (a) Faster RCNN^[1] (ResNet-50); (b) YOLOv3^[25] (Darknet-53); (c) FCOS^[32] (ResNet-50); (d) CenterNet^[24] (ResNet-18); (e) proposed method (ResNet-18)

Download full size

Fig. 8. False alarm rate of different approaches under different IoU threshold

Download full size

Stage	Sub-network	Block	Layer details				Output shape /(pixel×pixel×pixel)
Input	-	-	-		-		384×192×25
1	ABRRN	-	Deconv		Conv+ReLUConvRBAF-SSGAP		768×384×6
	ABRRN	-	Element-wise multiplicationSame as the above ABRRN				1536×768×3
2	ResNet-18backbone	Down sampling	Conv+BN+ReLU				384×192×64
		Down sampling	Max pooling				384×192×64
		ResBlock(no down sampling)	Conv+BN+ReLUConv+BN		Identity		384×192×64
		ResBlock(no down sampling)	Element-wise Addition				384×192×64
		ResBlock(down sampling)	Conv+BN+ReLUConv+BN		Conv+BN		192×96×128
		ResBlock(down sampling)	Element-wise addition				192×96×128
		ResBlock(no down sampling)	Same as the above ResBlock(no down sampling)				192×96×128
		ResBlock(down sampling)	Same as the above ResBlock(down sampling)				96×48×256
		ResBlock(no down sampling)	Same as the above ResBlock(no down sampling)				96×48×256
		ResBlock(down sampling)	Same as the above ResBlock(down sampling)				48×24×512
		Up sampling block	DCN+BN+ReLUDeconv+BN+ReLU				96×48×256
3	Up samplingnetwork	Up sampling block	Same as the above upsampling block				96×48×128
		Up sampling block	Same as the above upsampling block				96×48×64
4	RBOON	-	Conv+ReLU	Conv+ReLU		Conv+ReLU
			RBAF-SOD	Conv		Conv	96×48×1/2/2

Table 1. Overall structure of the detection network

Methods	AP50 /%			AP50:95 /%			Time /s
Methods	Small	Medium	All	Small	Medium	All	Time /s
Faster RCNN^[1](ResNet-50)	59.4	69.5	62.2	21.8	30.4	24.2	0.044
YOLOv3^[25](Darknet-53)	51.8	67.4	56.2	18.5	28.5	21.3	0.015
FCOS^[32](ResNet-50)	25.0	46.7	31.1	9.0	9.7	8.6	0.041

Table 2. Detection accuracy and ablation experiment

Methods	AP50 /%			AP50:95 /%			Time /s
Methods	Small	Medium	All	Small	Medium	All	Time /s
CenterNet^[24](ResNet-18)	60.2	70.4	63.1	21.5	28.3	22.7	0.011
Proposed-ABRRN(ResNet-18)	64.9	73.8	67.4	23.4	27.0	24.4	0.020
Ours-RBOON(ResNet-18)	60.7	70.6	63.5	21.6	28.4	22.9	0.011
Proposed(ResNet-18)	65.0	74.2	67.6	23.5	26.7	24.4	0.020

Table 3. Ablation experiment

Bofan Wang, Haitao Zhao. Small Object Detection in Hyperspectral Images Based on Radial Basis Activation Function[J]. Acta Optica Sinica, 2021, 41(23): 2311001

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information