DeepHawkeye is a library to detect unusual patterns in images using features from pretrained neural networks

Last update: Dec 13, 2022

Overview

English | 简体中文

Introduction

DeepHawkeye is a library to detect unusual patterns in images using features from pretrained neural networks

Reference PatchCore anomaly detection model

Major features

Using nominal (non-defective) example images only
Faiss(CPU/GPU)
TensorRT Deployment

Installation

$ git clone https://github.com/tbcvContributor/DeepHawkeye.git
$ pip install opencv-python
$ pip install scipy

# pytorch
$ pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html


#install faiss
# CPU-only version（currently available on Linux, OSX, and Windows）
$ conda install -c pytorch faiss-cpu
# GPU(+CPU) version （containing both CPU and GPU indices, is available on Linux systems）
$ conda install -c pytorch faiss-gpu
# or for a specific CUDA version
$ conda install -c pytorch faiss-gpu cudatoolkit=10.2 # for CUDA 10.2

Checkpoints and Demo data

Wide ResNet-50-2 and demo data

[Google]

[Baidu],code:a14e

${ROOT}
   └——————weights
           └——————wide_r50_2.pth
   └——————demo_data
           └——————grid
                    └——————normal_data
                    └——————test_data
           └——————....

Demo

bulid normal lib
python demo_train.py -d ./demo_data/grid/normal_data -c grid
pytorch infer
python demo_test.py -d ./demo_data/grid/test_data -c grid
tensorrt infer
python demo_trt.py -d ./demo_data/grid/test_data -c grid -t ./weights/w_res_50.trt

Tutorials

Need normal example images to cover all scenarios as much as possible
Faiss Documentation Default IVFXX, PQ16

train args

def get_train_args():
    parser = argparse.ArgumentParser()
    parser.add_argument('-d','--total_img_paths',type=str, default=None)
    parser.add_argument('-c','--category',type=str, default=None)
    parser.add_argument('--batch_size', default=64)
    parser.add_argument('--embedding_layers',choices=['1_2', '2_3'], default='2_3')
    parser.add_argument('--input_size', default=(224, 224))
    parser.add_argument('--weight_path', default='./weights/wide_r50_2.pth')
    parser.add_argument('--normal_feature_save_path', default=f"./index_lib")
    parser.add_argument('--model_device', default="cuda:0")
    parser.add_argument('--max_cluster_image_num', default=1000,help='depend on CPU memory, more than total images number')
    parser.add_argument('--index_build_device', default=-1,help='CPU:-1 ,GPU number eg: 0, 1, 2 (only on Linux)')

tips：

--input_size: trade off between speed and accuracy of the result --max_cluster_image_num：If RAM allows, greater than or equal to the total number of samples

test args

def get_test_args():
    parser = argparse.ArgumentParser()
    parser.add_argument('-d', '--test_path', type=str, default=None)
    parser.add_argument('-c', '--category', type=str, default=None)
    parser.add_argument('--model_device', default="cuda:0")
    parser.add_argument('--test_batch_size', default=64)
    parser.add_argument('--embedding_layers', choices=['1_2', '2_3'], default='2_3')
    parser.add_argument('--input_size', default=(224, 224))
    parser.add_argument('--test_GPU', default=-1, help='CPU:-1,'
                                                       'GPU: num eg: 0, 1, 2'
                                                       'multi_GPUs:[0,1,...]')
    parser.add_argument('--save_heat_map_image', default=True)
    parser.add_argument('--heatmap_save_path',
                        default=fr'./results', help='heatmap save path')
    parser.add_argument('--threshold', default=2)
    parser.add_argument('--nprobe', default=10)
    parser.add_argument('--n_neighbors', type=int, default=5)
    parser.add_argument('--weight_path', default='./weights/wide_r50_2.pth')
    parser.add_argument('--normal_feature_save_path', default=f"./index_lib")

tips：

--threshold: depend on scores of anomaly data

result format：{filename}_{score}.jpg

TensorRT Deployment

License

This project is released under the Apache 2.0 license.

Code Reference

https://github.com/hcw-00/PatchCore_anomaly_detection embedding concat function : https://github.com/xiahaifeng1995/PaDiM-Anomaly-Detection-Localization-master

DeepHawkeye is a library to detect unusual patterns in images using features from pretrained neural networks

Related tags

Overview

Introduction

Installation

Checkpoints and Demo data

Demo

Tutorials

License

Code Reference

Owner

CV Newbie

A Rao-Blackwellized Particle Filter for 6D Object Pose Tracking

Several simple examples for popular neural network toolkits calling custom CUDA operators.

Learning Saliency Propagation for Semi-supervised Instance Segmentation

ImageNet-CoG is a benchmark for concept generalization. It provides a full evaluation framework for pre-trained visual representations which measure how well they generalize to unseen concepts.

EMNLP 2020 - Summarizing Text on Any Aspects

Multimodal Descriptions of Social Concepts: Automatic Modeling and Detection of (Highly Abstract) Social Concepts evoked by Art Images

Code for the paper "Improving Vision-and-Language Navigation with Image-Text Pairs from the Web" (ECCV 2020)

Deep Learning Based Fasion Recommendation System for Ecommerce

An SE(3)-invariant autoencoder for generating the periodic structure of materials

Simulation environments for the CrazyFlie quadrotor: Used for Reinforcement Learning and Sim-to-Real Transfer

Source Code of NeurIPS21 paper: Recognizing Vector Graphics without Rasterization

CS_Final_Metal_surface_detection - This is a final project for CoderSchool Machine Learning bootcamp on 29/12/2021.

Auxiliary data to the CHIIR paper Searching to Learn with Instructional Scaffolding

DM-ACME compatible implementation of the Arm26 environment from Mujoco

Cross-modal Retrieval using Transformer Encoder Reasoning Networks (TERN). With use of Metric Learning and FAISS for fast similarity search on GPU

Pre-trained NFNets with 99% of the accuracy of the official paper

Learning to Adapt Structured Output Space for Semantic Segmentation, CVPR 2018 (spotlight)

UMEC: Unified Model and Embedding Compression for Efficient Recommendation Systems

《Deep Single Portrait Image Relighting》(ICCV 2019)

An open-source Deep Learning Engine for Healthcare that aims to treat & prevent major diseases