Introduction - Gachon-Graduation-work/Muzzle_Detection_Project GitHub Wiki

Motivation

In recent years, various dog bites have occurred. Such dog bites occur frequently, reaching an average of 2030 cases per year. Although there are laws related to such dog bites, it was judged that the currently enforced laws were ineffective regulations because they were not suitable for accident prevention.

So we implemented this project to solve these problems.

This project allows users to identify whether a dog is a fierce dog by taking pictures of it.

Introduction

The Muzzle Detection project is designed to classify whether the dog is a fierce dog when the user takes a picture of the dog on the user's smart phone, and to automatically check whether the dog is wearing a muzzle if it is a fierce dog, and to inform the results. In this project, one Android device is used as client(App-Camera) and one PC is used as Python Server to run model.

When the user starts our app and presses the photo button, the camera function is executed, allowing the user to take pictures with the camera of the mobile phone. When you press the Take button, the app automatically converts the photo into bytes and sends it to a Python server operated by a PC.

The Python server that receives the picture converts the byte back into an image. Then, the photo enters the input data in the existing Yolo model and checks whether a dog is present in the image. If a dog exists, only the part where the dog exists is cropped and stored as a separate photo.

The images of each dog go into the input data of the muzzle detection Yolo model we created and determine whether muzzle exists in the image. If a muzzle is found, the dog determines that it is safe and stops the process. However, if the muzzle is not found, the image of the dog enters the input data in the resnet dog classification model we created to determine whether the dog is a blind dog or not. If the dog is not a fierce dog, the dog is judged to be safe, but if it is fierce dog and it is not wearing muzzle, then it is judged to be dangerous.

Each dog's photo will have determined information (whether to wear a muzzle or a fierce dog), and at the end of the entire process, an image with information on the area of each dog's cropped image will be overwritten in the photo initially sent to the server.

The resulting image is converted back into bytes and transmitted to the user. After receiving the photo, the app prints it out to the user and informs them of the results.