VisionSphere — Explainable Image Recognition Platform

# VisionSphere: Explainable Computer Vision During my Python Developer Internship at **Codec Technologies Pvt. Ltd.** (September–October 2026), I was tasked with engineering an image classification application that went beyond returning a single label. Standard classification systems operate as black boxes; VisionSphere was designed to make deep learning predictions explainable and inspectable. ## The Core Challenge Deep learning models can fail in non-obvious ways. A neural network might classify a picture of an animal correctly, but for entirely spurious reasons (such as recognizing background snow instead of a wolf). VisionSphere tackles this by integrating **Grad-CAM (Gradient-weighted Class Activation Mapping)** to generate heatmaps highlighting the exact pixel clusters that triggered the output class. ## Architecture & Technical Execution 1. **Multi-Architecture Model Pool**: Integrated pretrained MobileNetV3 (lightweight, edge-friendly), ResNet50 (robust 50-layer residual network), and EfficientNet-B0 architectures. 2. **OpenCV Preprocessing Pipeline**: Built filters for brightness compensation, Gaussian blur noise reduction, and Canny edge detection. 3. **Heatmap Generation**: Computed gradients of the score for target class $c$ with respect to feature activation map $A^k$ of the final convolutional layer. 4. **Flask Interface**: Deployed as an interactive web dashboard with drag-and-drop file ingestion, live webcam diagnostics, and JSON API endpoints.