Image Processing Projects in Bangalore

A U-Net architecture is trained on the BraTS 2020 dataset containing 369 multi-modal MRI scans with expert-annotated tumour segmentation masks. The model is trained with a combined Dice and binary cross-entropy loss. Post-processing with connected-component analysis removes false-positive regions smaller than 500 voxels. The model achieves a mean Dice score of 0.87 on the validation set for whole-tumour segmentation. A web application allows radiologists to upload a DICOM file and view the predicted tumour mask overlaid on the MRI slice with a volume estimate in cubic centimetres.

A face recognition attendance system captures video from a webcam, detects faces using MTCNN, and computes 128-dimensional embeddings using a pre-trained FaceNet model. Student embeddings are enrolled in a FAISS index during registration. At attendance time, detected face embeddings are matched against the index with a cosine distance threshold of 0.4. Recognised students are marked present in a SQLite database. The system processes 30 FPS video and correctly identifies 98.3% of enrolled students in a 50-person test under varied lighting conditions. A Flask dashboard displays attendance records with timestamps.

A convolutional neural network is trained on the DeepPCB dataset containing 1,500 PCB image pairs with six defect types: open circuit, short circuit, mouse bite, spur, copper protrusion, and missing hole. ResNet-50 fine-tuned with data augmentation achieves 96.8% classification accuracy and 94.2% defect localisation mAP. The inference pipeline processes a PCB image in 120 ms on a CPU. A Tkinter desktop application allows quality control engineers to load a PCB image and receive an annotated output highlighting defect locations with type labels and confidence scores.

A document scanning application uses OpenCV to detect document boundaries via contour detection and perspective transform, corrects skew and illumination, and passes the rectified image to Tesseract OCR for text extraction. The pipeline handles A4 documents, business cards, and handwritten notes. Post-processing with a spell-checker and layout analysis separates the extracted text into paragraphs, tables, and headers. A mobile-friendly React web app allows users to upload a photo of a document and download the extracted text as a formatted PDF. Processing a standard A4 page takes 2.3 seconds end to end.

A transfer learning model based on EfficientNet-B3 is trained on the PlantVillage dataset containing 54,306 images of healthy and diseased plant leaves across 38 classes. Data augmentation with random crops, colour jitter, and horizontal flips is applied during training. The model achieves 98.1% top-1 accuracy on the test set. Grad-CAM visualisations confirm that the model attends to lesion regions rather than background artefacts. A React Native mobile app allows farmers to photograph a leaf and receive an instant disease diagnosis with treatment recommendations sourced from a curated knowledge base.

An automatic number plate recognition system processes video from a parking lot camera. YOLOv8 detects number plate regions; a custom CNN trained on 10,000 Indian number plate images performs character segmentation and recognition. The system achieves 94.7% plate-level recognition accuracy on a test set of 500 frames under varied lighting and angle conditions. Recognised plates are matched against a registered vehicle database; unregistered vehicles trigger an alert. A Node.js backend stores entry/exit timestamps and computes parking duration for billing. The system processes video at 15 FPS on a Raspberry Pi 4.

A crowd analysis system uses YOLOv8 for pedestrian detection and a perspective transform to map image coordinates to real-world distances. Social distancing violations are flagged when the estimated distance between two people falls below 1.5 metres. Crowd density is estimated using a CSRNet model trained on the ShanghaiTech dataset, achieving a mean absolute error of 9.8 people per frame. The system processes CCTV footage at 20 FPS and generates a heatmap of high-density zones. A Flask dashboard displays live violation counts, density heatmaps, and hourly crowd trend charts.

A convolutional neural network is trained on the MNIST dataset for digit recognition and extended to the EMNIST Balanced dataset for 47-class alphanumeric character recognition. The architecture uses three convolutional blocks with batch normalisation and dropout, achieving 99.4% accuracy on MNIST and 87.6% on EMNIST Balanced. A drawing canvas built with HTML5 Canvas and React allows users to write digits or characters with a mouse or touchscreen and receive real-time predictions with a confidence bar chart. The model is exported to ONNX format and runs in the browser via ONNX Runtime Web without a server round-trip.

A deep learning pipeline grades diabetic retinopathy severity from retinal fundus photographs using the Kaggle DR Detection dataset of 35,126 images across five severity grades. A DenseNet-121 backbone with a custom classification head is trained with label smoothing and mixup augmentation. The model achieves a quadratic weighted kappa of 0.83 on the validation set, matching published clinical benchmarks. Grad-CAM overlays highlight microaneurysms and haemorrhages driving the prediction. A web application allows ophthalmologists to upload a fundus image and receive a grading report with annotated regions of concern.