Method for Classifying Crime Scene Photographs Based on Convolution Neural Network

Zhuorong Li; Yunqi Tang; Nengbin Cai

doi:10.3788/LOP212827

Laser & Optoelectronics Progress, Volume. 60, Issue 4, 0410007(2023)

Method for Classifying Crime Scene Photographs Based on Convolution Neural Network

Zhuorong Li¹, Yunqi Tang^1、*, and Nengbin Cai²

Author Affiliations

¹School of Criminal Investigation, People's Public Security University of China, Beijing 100038, China

²Shanghai Key Laboratory of Crime Scene Evidence, Shanghai 200083, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(18)

Fig. 1. Display of dataset. (a) Orientation; (b) outline; (c) key part

Download full size

Fig. 2. Crime scene photos. (a) Outline of the hotel lobby; (b) outline of the dormitory; (c) outline of the convenience store; (d) key part

Download full size

Fig. 3. Structure of CriSNet

Download full size

Fig. 4. Specific structure of optimization options. (a) Option M1-1; (b) Option M1-2; (c) Option M2-1; (d) Option M2-2

Download full size

Fig. 5. Three different kinds of residual block structures. (a) Original residual blocks; (b) pre-activated residual blocks optimized by Ref. [30]; (c) proposed pre-activated residual blocks in this paper

Download full size

Fig. 6. Comparison of baseline network model

Download full size

Fig. 7. Confusion matrix of CriSNet

Download full size

Fig. 8. ROC of network model

Download full size

Fig. 9. Distribution histogram of misclassification confidence

Download full size

Fig. 10. Representative misclassification pictures. (a) (b) (c) Outline; (d) key part

Download full size

Table 1. Parameters of CriSNet

View table

Table 1. Parameters of CriSNet

Type	Number	Kernel size /stride		Output size
Convolution/BN/Scale	1	7×7/S2		［64，112，112］
Max-pooling	1	3×3/S2		［64，56，56］
Bottleneck	1	1×1/S1 3×3/S1 1×1/S1	1×1/S1	［256，56，56］
Inception a	2	1×1/S1 3×3/S1 5×5/S1 1×1/S1	1×1/S1 1×1/S1 3×3（max-pooling）	［480，56，56］
Max-pooling	1	3×3/S2		［480，28，28］
Inception b	5	1×1/S1 3×3/S1 5×5/S1 1×1/S1	1×1/S1 1×1/S1 3×3（max-pooling）	［832，32，32］
Max-pooling	1	3×3/S2		［832，14，14］
Inception c	2	1×1/S1 3×3/S1 5×5/S1 1×1/S1	1×1/S1 1×1/S1 3×3（max-pooling）	［1024，14，14］
Ave-pooling	1	7×7/S1		［1024，8，8］
Innerproduct	1	-		［1，4］

Table 2. Composition of the dataset
View table
Table 2. Composition of the dataset
Classification Orientation Outline Key part Negative sample
Training set 2926 3756 3832 3208
Verification set 400 450 500 400
Test set 350 450 500 400
Summary 3676 4656 4832 4008

Table 3. Results of different optimization schemes

View table

Table 3. Results of different optimization schemes

Scheme	Method	Time /min	Accuracy /%	F1 Score /%
Scheme（a）：Conv/LRN	Option M1-1	61	93.00	92.87
Scheme（b）：Conv/BN	Option M1-2	57	93.53	93.11
Scheme（c）：3×3 Conv/S2	Option M1-2+Option M2-1	28	93.32	92.96
Scheme（d）：proposed	Option M1-2+Option M2-2	52	93.83	94.16
Ref.［30］	-	61	92.86	92.72

Table 4. Comporision of the baseline network model

View table

Table 4. Comporision of the baseline network model

Parameter	AlexNet /%	VGGNet /%	GoogLeNet /%	ResNet /%	MobileNet /%	EfficientNet /%	CriSNet /%
Accuracy	91.53	90.82	92.06	92.34	91.34	92.75	93.83
Precision	91.88	91.88	92.51	92.46	92.26	93.14	94.01
Recall	92.46	91.70	92.97	93.36	92.45	92.72	94.38
F1 Score	92.12	91.75	92.70	92.77	92.34	92.92	94.16

Table 5. Distribution of misclassification confidence

View table

Table 5. Distribution of misclassification confidence

Scores	AlexNet /%	VGGNet /%	GoogLeNet /%	ResNet /%	MobileNet /%	EfficientNet /%	CriSNet /%
>0.9	63.89	63.46	36.30	29.55	23.13	30.50	21.29
0.8-0.9	09.72	12.18	16.30	21.97	17.16	14.40	21.29
0.7-0.8	09.03	11.54	20.74	12.88	17.91	21.18	16.67
0.6-0.7	07.64	05.77	08.15	16.67	19.40	17.79	19.44
<0.6	09.72	07.05	18.52	18.94	22.38	16.10	21.29

Table 6. Influence of the resolution of training set on the accuracy of network

View table

Table 6. Influence of the resolution of training set on the accuracy of network

Resolution	Accuracy /%
Resolution	AlexNet	VGGNet	GoogLeNet	ResNet	MobileNet	EfficientNet	CriSNet
48×48	84.40	89.83	88.51	88.86	78.53	88.75	91.03
96×96	89.20	92.06	90.26	91.60	85.04	90.98	92.57
224×224	91.53	90.82	92.06	92.34	91.43	92.75	93.83

Table 7. Influence of noise on the accuracy of network
View table
Table 7. Influence of noise on the accuracy of network
Noise density Accuracy /%
AlexNet VGGNet GoogLeNet ResNet MobileNet EfficientNet CriSNet
0.01 91.09 91.14 92.11 91.43 91.50 92.50 93.60
0.10 88.69 90.17 91.03 91.20 90.25 92.00 93.60
0.50 76.00 75.83 76.34 82.29 76.25 86.50 84.63

Table 8. Confidence of misclassification pictures

View table

Table 8. Confidence of misclassification pictures

Network	Picture	Ground truth	Predict label	Score
AlexNet	Fig. 10（a）	Outline	Outline	0.9967
	Fig. 10（b）	Outline	Outline	0.9948
	Fig. 10（c）	Outline	Key part	0.8179
	Fig. 10（d）	Overview	Orientation	0.9999
VGGNet	Fig. 10（a）	Outline	Outline	0.9954
	Fig. 10（b）	Outline	Outline	0.5684
	Fig. 10（c）	Outline	Key part	0.7912
	Fig. 10（d）	Overview	Outline	0.9999
GoogLeNet	Fig. 10（a）	Outline	Outline	0.9954
	Fig. 10（b）	Outline	Outline	0.5684
	Fig. 10（c）	Outline	Outline	0.9916
	Fig. 10（d）	Overview	Orientation	0.9995
ResNet	Fig. 10（a）	Outline	Outline	0.9037
	Fig. 10（b）	Outline	Orientation	0.6594
	Fig. 10（c）	Outline	Key part	0.7885
	Fig. 10（d）	Overview	Outline	0.8860
MobileNet	Fig. 10（a）	Outline	Outline	0.9934
	Fig. 10（b）	Outline	Key part	0.7860
	Fig. 10（c）	Outline	Outline	0.9975
	Fig. 10（d）	Overview	Orientation	0.9907
EfficientNet	Fig. 10（a）	Outline	Key part	0.8836
	Fig. 10（b）	Outline	Key part	0.8839
	Fig. 10（c）	Outline	Key part	0.8799
	Fig. 10（d）	Overview	Key part	0.8851
CriSNet	Fig. 10（a）	Outline	Outline	0.9728
	Fig. 10（b）	Outline	Outline	0.9941
	Fig. 10（c）	Outline	Outline	0.9287
	Fig. 10（d）	Overview	Orientation	0.9881

Tools

Get Citation

Copy Citation Text

Zhuorong Li, Yunqi Tang, Nengbin Cai. Method for Classifying Crime Scene Photographs Based on Convolution Neural Network[J]. Laser & Optoelectronics Progress, 2023, 60(4): 0410007

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites