Akshay Dabholkar Bhushan Muthiyan Shilpa Srinivasan Swetha Ravi Hyeran Jeon Jerry Gao Department of Computer Engineering San Jose State University San Jose California ID: 628051
Download Presentation The PPT/PDF document "Smart Illegal Dumping Detection" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Smart Illegal Dumping Detection
Akshay
Dabholkar, Bhushan Muthiyan, Shilpa Srinivasan, Swetha Ravi, Hyeran Jeon, Jerry GaoDepartment of Computer EngineeringSan Jose State UniversitySan Jose, CaliforniaSlide2
Illegal Dumping
Illegal dumping has been a chronic problem in many cities in the world.
Illegal dumping makes our communities unsafe and it is expensive for the city to remove those items.
To reduce the illegal dumping, a few cities have designed community-based voluntary reporting systems and surveillance camera-based monitoring systems.Slide3
Existing Work
Los Angeles Health and
Sanitation
Chicago Clean Drive
Breathe Project PittsburghAir LouisvilleBeautify San Jose
Ref :http://datasmart.ash.harvard.edu/news/article/monitoring-air-quality-and-the-impacts-of-pollution-679
https://www.airlouisville.com/Slide4
SpotGarbage
Mobile application based solution to tackle garbage litter in India.
Key difference between our approach and theirs is on the image dataset and litter they cater to.
Ref :https://github.com/
KudaP/SpotGarbageSlide5
Our ApproachPropose to use deep learning technique to recognize various types of frequently dumped wastes with high recognition accuracy and small memory footprint.Slide6
Our GoalSlide7
Technologies UsedSlide8
Deep Learning - The new trend setter
Deep learning is a multilayered, back propagating, representational, self-learning method implemented by using simple but non-linear functions.Deep learning is the part of the machine learning based on the representational learning methodsSlide9
Neural network
The fundamental element of the deep learning or machine learning is the artificial neuron.
Calculations for learning are performed by these neurons.
Different types are
Feed ForwardRadial Basis FunctionMulti Layer PerceptronRecurrent Neural NetworkSelf Organizing MapAnd many more…..
Ref : http://cs231n.github.io/neural-networks-1/ https://www.extremetech.com/extreme/215170-artificial-neural-networks-are-changing-the-world-what-are-theySlide10
Convolutional Neural Network (CNN)
Convolutional Neural Network (CNN) is a practically useful algorithm especially in computer vision tasks by reduced complexity with resolving overfitting problems from deep learning.
CNN inside is made of simple repeated matrix multiplications without branch operations.
Ref:https
://devblogs.nvidia.com/parallelforall/deep-learning-nutshell-core-concepts/Slide11
Neural Network Frameworks
Ref: http://blog.algorithmia.com/deploying-deep-learning-cloud-services/Slide12
MethodologyCreate illegally dumped material labeled dataset.
Create lmdb database from the dataset.Tweak the existing model as per need and train the model using Caffe interface with the generated
lmdb files.Deploy the model file on the embedded platform Jetson TX1Test the model for prediction accuracy.Slide13
Illegal dump object classification
We classified the illegal dumped objects into following broad category :
Chair
MattressTableSofaFurnitureTrashSlide14
Proposed approaches
Approach 1 (Naive Approach)Frequently dumped wastes identified in 102 illegal dumping images provided by City of San Jose
6 classes
Baseline Network: AlexNet
Iterations: 100Image dataset size:161 MbSlide15
Prediction Accuracy of Approach 1 Slide16
This was insufficient…
Solution : Add more categories and train for Clean Area detection.Slide17
Increased Classification
Electronics
Trees
Cart
Clean AreaSlide18
Approach 2
Approach with more classes: 11 classes
Baseline Network: AlexNet and
GooglenetIterations: 5000
Image dataset size: 802 MbSlide19
Prediction Accuracy of Approach 2Slide20
Misprediction
The disadvantage of this approach was the model tends to predict clean area even for the image with illegal dump object because the feature map for clean areas was more compared to normal images.Slide21
Revise Approach
Solution Image pre-processing to clearly define the region of interest.
Excluding Clean Area classification.Classes AggregationSlide22
Approach 3
Approach with pre-processed imagesCropping image dataset for better training.
8 classes
Baseline Network: AlexNet and Googlenet
Iterations: 10000Image dataset size: 1.14 GbSlide23
Prediction Accuracy of Approach 3Slide24
Energy Efficiency Approach- TensorRT
We used NVIDIA TensorRT that helps shrink model size.
Uses two approaches to shrink trained model size. 1. Quantization, 2. Optimizations by applying vertical and horizontal layer fusion.
Ref: https://developer.nvidia.com/tensorrtSlide25
Model size variation using TensorRT
Approaches
Network
Original (in MB)
With
TensorRT
(in MB)
Approach 1
AlexNet
227.6
84.7
Approach 2
AlexNet
227.7
113.8
GoogleNet
41.4
12
Approach 3
AlexNet
227.6
113.8
GoogleNet
41.3
12Slide26
ConclusionDetect illegal dumping with over 90% accuracy and reduced memory consumption.
The model could be scaled to cover more classifications by training with more images of greater variety.Extensive pre-processing could be done to cover the adverse lighting conditions and thereby increase prediction accuracy.Slide27Slide28
Thank youSlide29
Why Caffe?
Caffe is a deep learning framework made with expression, speed, and modularity in mind. It is developed by the Berkeley Vision and Learning Center (BVLC
) and by community contributors.Expressive architecture
Extensible codeSpeed
Ref : http://caffe.berkeleyvision.org/Slide30
We have chosen Caffe as our baseline framework as it works
well for image recognition purposes.
Ref: https://www.slideshare.net/NiketanPansare/notes-from-2016-bay-area-deep-learning-schoolSlide31
Jetson TX1Slide32
References
[1] “About drive clean
chicago,” 2017. [Online].
Available:http://www.drivecleanchicago.com/About/What.aspx
[2] “#beautifysj,” 2017. [Online]. Available: http://www.beautifysj.org/[3] “Los angeles applications,” 2017. [Online]. Available: http://geohub.
lacity.org/
[4] “Streets
philadelphia
,” 2017. [Online]. Available: http://www.philadelphiastreets.com/philly-spring-cleanup/
[5] . A. R. M.
Ballester
, P., “On the performance of
googlenet
and
alexnet
applied to sketches,” The Thirtieth AAAI Conference on Artificial Intelligence,2016.
[6] I. G. Y.
Bengio
and A.
Courville
, “Deep learning,” 2016, book in preparation for MIT Press. Available:
http://www.deeplearningbook.org
.
[7] Y. J. P. S. S. R. D. A. D. E. V. V. A. R. Christian
Szegedy
, Wei Liu, “Going deeper with convolutions,” IEEE Conference on Computer Vision and Pattern Recognition, 2015.
[8] Y. Jia, E.
Shelhamer
, J. Donahue, S.
Karayev
, J. Long, R.
Girshick,S
.
Guadarrama
, and T. Darrell, “Caffe: Convolutional architecture for fast feature embedding,” in Proceedings of the 22Nd ACM International Conference on Multimedia, 2014, pp. 675–678.[9] Y. LeCun, B. Boser
, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel, “Backpropagation applied to handwritten zip code recognition,” Neural
Comput., vol. 1, no. 4, pp. 541–551, 1989.[10] A. Moujahid, “A practical introduction to deep learning with
caffe and python,” 2017. Available: http://adilmoujahid.com/posts/2016/06/introduction-deep-learning-python-caffe/
[11] NVIDIA, “Nvidia tensorrt,” 2017. Available:
https://developer.nvidia.
com/
tensorrt
[12] J.
Salleh
and M.
Tsudagawa
, “Classification of industrial disposal illegal dumping site images by using spatial and spectral information together,” IEEE Instrumentation and Measurement Technology Conference, 2002.