Eliminate image squeezing by forcing square cropping in both preprocessing and browser bypass paths

Fix model loading by downloading snapshot and renaming .pth weight file

Shift classification model to ResNet-50 ConvNet architecture with Softmax activation and French-to-English translation

Fix target crop coordinate offset using a temporary canvas to guarantee perfect alignment

Restore confidence threshold check to 75% per user request

Use sigmoid activation on SigLIP2 logits to match sigmoid loss training

Add /api/debug_crop endpoint to verify AI input frames

Fix missing glass category mapping and lower confidence threshold to 40% for SigLIP2 accuracy

Implement stable center target reticle to fix handheld camera tracking jitters

Completely clean up and remove all legacy models and YOLO references, leaving SigLIP2 as the sole classifier

Bypass server-side cropping for browser-uploaded frames to prevent double cropping

Update transformers dependency to force latest version for SigLIP2 compatibility

Optimize classification latency and fix canvas coordinate cropping scaling bug

Integrate YOLOv8 for deep-learning-based object detection and crop

Restore original dark blue/violet theme and re-integrate features

Update site UI theme to a premium high-tech dark green cyber-organic design

Update site UI theme to match project presentation slide aesthetics

Optimize responsive design layout for all device sizes

Fix smartphone aspect ratio collapse and clean head scripts

Move device camera scanner to a sliding tab interface

Deploy waste classifier to Hugging Face Spaces (clean binaries)