Face recognition is a key computer vision technology that is widely used in identity verification, security monitoring, and social media. The model implementation is divided into two parts, one is the image recognition and cutting of yolo method, the other is the code part and the model implementation part. The project combines feature extraction methods such as PCA, NMF, and CNN to analyze existing schemes and propose an optimization scheme that combines traditional methods and deep learning to improve the performance and flexibility of face recognition systems.

Week 1: （2025/2/2）

This week, we held a panel discussion on the project and confirmed the subject model of "face recognition optimization scheme". After each member understood the scope of knowledge, relevant technology and tool resources needed to carry out the project, we identified and collected all the components needed for the project. And assign the tasks and responsibilities of each member of the project team to the project.

Aim for week 1:

Debugging, selection, configuration and implementation of YOLO face detection model; Develop and optimize the face recognition model, and gradually improve the recognition performance of the model

Implementation of tasks:

We divided the team of six into two parts. One part is responsible for debugging, selecting, getting familiar with and configuring YOLO, and the other part is responsible for optimizing the face recognition model through various tools and code.

FulinYang : 1, confirmed what is YOLO face detection; 2, the benefits of choosing YOLO for face detection; 3, the method of training YOLOv5 for face detection; 4, confirmed the python code needed for inference (face detection). 5. Implementation method of YOLO (divided into data set selection, data annotation format and training process)

Siqi Jia, Jinrun Tan: 1, confirming the best YOLO version (YOLOv8) for completing the model. 2, explains why YOLOv8 is the best version to complete the project. 3. Configure the steps for setting YOLOv8. 4. Complete the installation steps and initial operation of YOLOv8.

Peng Wang: 1, developed a facial recognition optimization system combining traditional feature extraction methods and deep learning models. 2, use the code from the scikit-learn website example for a preliminary test. 3. NMF technology was introduced to optimize the model in combination with PCA and SVM. 4. Performance evaluation after model optimization

Ziyu Huang, Yicheng Sheng: 1. Use CNN (Convolutional neural network) to replace NMF for face recognition and complete the secondary optimization of the model. 2. Solve the misjudgment problem of Donald Rumsfeld and Gerhard Schroeder, and improve the recognition rate of Hugo Chavez categories; 3. Integrate data enhancement technology and introduce FaceNet for feature extraction

Issues and solutions:

Issue 1：

Small face detection:

Solution：

Appropriately increase the input resolution or multi-scale training, adding more small face samples.

Issue 2：

Too many false checks:

Solution：

Adjust the confidence threshold, or optimize the data quality.

Issue 3：

Lack of speed:

Solution：

Use lightweight models (like YOLOv5s or n versions), or speed up inference on Gpus/TensorRT/OpenVINO.)

Week 2: （2025/2/9）

This week, through division of labor and collaboration, we have completed the preliminary coding of the YOLO v11 model, and the model can more accurately identify and segment objects in images.

Aim for week 2:

Debugging, configuration and implementation of YOLO face detection model; Develop and optimize the face recognition model, and gradually improve the recognition performance of the model; Introduce automated scripts

Implementation of tasks:

Siqi Jia, Jinrun Tan: 1. Introduce automated scripts to handle the time-consuming processing of documents in multiple language versions to ensure consistency and accuracy of the documents. 2. Test model export, confirm callback information, and export functionality is normal. 3. Detect image segmentation effects and use segmentation models to train COCO segmentation datasets

FulinYang : 1. The method of early stopping and learning rate self modification was used to reduce the overfitting of the YOLO V11 model. 2. The official dataset of Ultralytics was used for iterative training, and the final model had 25 iterations 

Ziyu Huang, Yicheng Sheng: 1. Hard Negative Mining: Identified and retrained on 38 hard negative samples, allowing the model to handle challenging misclassifications more effectively. 2.Feature Extraction with InceptionResNetV2: Enhanced feature representation by using a more sophisticated deep learning architecture. 3.Optimized Data Augmentation: Improved training data diversity with controlled transformations. 4.Hyperparameter Fine-Tuning: Adjusted SVM parameters to maximize model generalization. 5. Integrate Hard Negative Mining into our pipeline, improve the model’s recall, accuracy, and overall robustness

Peng Wang: The preliminary coding of the YOLO v11 model has been completed and tested in the local environment. The README package now includes a quick start guide and installation steps. A documentation site has been set up in MkDocs, and complete project documentation can now be generated and viewed. 3. All major functional unit tests have been passed, and integration testing has revealed performance bottlenecks in high concurrency situations, which have been optimized. 4. The model has been successfully deployed to the cloud platform, containerized using Docker, and a daily maintenance and monitoring plan has been developed to ensure model performance and stability.

Issues and solutions:

Issue 1：

underperforming classes:

Solution：

integrating Hard Negative Mining into  pipeline

Issue 2：

In the case of multiple language versions, manually building and updating documents for document projects can be time-consuming 

Solution：

Construct automated script construction in MkDocs project

Week 3: （2025/2/16）

This week, based on the test1.py obtained in the previous two weeks, we attempted to further improve the YOLO and SVM models to enhance the accuracy of facial recognition

Aim for week 3:

Based on test1.py, the original face recognition model is enhanced by integrating FaceNet and SVM

 Implementation of tasks:

1. Integrate FaceNet and SVM to enhance the original face recognition model. 2 Download and configure YOLOv111 runtime environment 3 from Ultralytics' official GitHub 200 images were selected from a dataset of 1288 facial images, and five different YOLOv11 models were used to evaluate the facial detection capability of YOLOv11 The result is not satisfactory, YOLOv11 is difficult to accurately detect faces

Problems and possible causes:

Question 1:

YOLOv11 struggled to accurately detect faces, regardless of whether using lighter or more accurate models. The detection confidence is usually too low, and many images cannot record any detected faces at all

Possible reasons:

1. Low detection accuracy: Even with the most advanced model YOLO11x, face detection is still unreliable.

2. Possible dataset mismatch: The LFW dataset may not match the training data of YOLOv11, resulting in poor performance.