Overview / Usage

• OverView
  Learning Deep Features for One-Class Classification (AnomalyDetection).
  Corresponds RaspberryPi3 and LaptopPC.
  Convert to Tensorflow, ONNX, Caffe, PyTorch.
  This project was inspired by Image abnormality detection using deep learning ーPapers and implementationー - Qiita - shinmura0, Image inspection machine for people trying hard - Qiita - shinmura0 and was created.
  I would like to express my deepest gratitude for having pleasantly accepted his skill, consideration and article quotation.
  His articles that were supposed to be used practically, not limited to logic alone, are wonderful.
  However, I don't have the skills to read papers, nor do I have skills to read mathematical expressions.
  I only want to verify the effectiveness of his wonderful article content in a practical range.
  To be honest, I am not engaged in the work of making a program.
  There are many methods such as methods using "Implemented ALOCC for detecting anomalies by deep learning (GAN) - Qiia - kzkadc" and methods using "Detection of Video Anomalies Using Convolutional Autoencoders and One-Class Support Vector Machines (AutoEncoder)" for image anomaly detection using deep learning.
  Here is an article on detecting abnormality of images using "Variational Autoencoder".
  Image abnormality detection using Variational Autoencoder (Variational Autoencoder) - Qiita - shinmura0
  The method to be introduced this time is to detect abnormality by devising the loss function using normal convolution neural network(CNN).
  In conclusion, it was found that this method has good anomaly detection accuracy and visualization of abnormal spots is also possible.
  　
• Usage
  https://github.com/PINTO0309/Keras-OneClassAnomalyDetection#10-implementation-by-raspberrypi
  　
• Future tasks
  • Cooperation with RealSense D435
    　
• Issue
  RuntimeError: AssertionFailed: isOrdersCompatible(_dimsOrder, dimsOrder)
  How to offload OpenVINO non-compliant layer to Tensorflow (undefined symbol: _ZN15InferenceEngine10TensorDescC1Ev)

Methodology / Approach

• 「Learning Deep Features for One-Class Classification」 (Subsequent abbreviations, DOC) arxiv： https://arxiv.org/abs/1801.05365
• Cost is $100 or less (conventional products have over $9,000)
• Absolute detection accuracy is the highest peak (state-of-the-art at the time of publication)
• Compact (RaspberryPi and Web Camera only)
• Despite using deep learning at RaspberryPI it is fast (5 FPS - 15 FPS)
• Application areas
  • Visual inspection of industrial products
  • Appearance inspection of bridges by drone here
  • Surveillance camera here

Technologies Used

• DOC
• OpenVINO
• Tensorflow Lite
• NCS/NCS2
• Intel HD Graphics (GPU)

Repository

https://github.com/PINTO0309/Keras-OneClassAnomalyDetection

Other links

• Image inspection machine for people trying hard - Qiita - shinmura0
• Image abnormality detection using deep learning ーPapers and implementationー - Qiita - shinmura0
• [180 FPS] High performance tuning of DOC