Animal Detection Neural Network
A neural network system that segments camouflaged animals from their natural environments using multi-class semantic segmentation.
Gallery
About This Project
AnimalDetection is an academic deep learning project from a Biologically Inspired Artificial Intelligence course. It trains a UNet-based segmentation model on a custom ~300-image dataset to locate and delineate camouflaged animals — scenarios where standard object detection fails because the target blends into its background.
Rather than binary foreground/background segmentation, the system decomposes each scene into four semantic regions: the animal, the masking background (visually similar to the animal), the non-masking background, and a foreground attention region. This four-class formulation is better suited to understanding why a region is difficult to segment. Final IoU: 0.51 on a custom dataset, trained on consumer hardware (NVIDIA GTX 1070, 8 GB VRAM).
Features
- Four-class segmentation — animal / masking background / non-masking background / foreground attention
- UNet + ImageNet pretraining — ResNet34 encoder pretrained on ImageNet; decoder path fine-tuned on ~300 camouflage images
- Weighted cross-entropy loss — class weights
[1, 0.05, 0.03, 0.01]strongly prioritize correct animal-pixel prediction - Multi-group stochastic augmentation — three independent
OneOfgroups targeting tone, blur, and signal degradation axes - Timestamped training runs — each run saves checkpoint, config dump, stats, live loss/IoU plot, and composite visualizations
Technical Architecture
Four Python packages: config (hyperparameters, class definitions, model/loss/optimizer), data_model (custom Dataset, train/valid/test splits, epoch wrappers), utils (augmentation, preprocessing, visualization, file I/O), and a top-level main.py loop.
Data flows: disk → OpenCV (BGR→RGB) → dimension crop to 32-pixel multiples (UNet architectural constraint) → color-based mask extraction into (H, W, 4) float array → albumentations augmentation → normalization + transposition to (C, H, W) tensor.
Ground-truth masks are stored as RGB color images where each pixel’s color encodes its semantic class. At load time, cv2.inRange performs exact-color pixel matching per class; results are normalized to [0, 1] float and stacked channel-wise into the multi-class target tensor.
Engineering Highlights
Transfer learning on small data — With ~300 images, training from scratch would underfit. ImageNet-pretrained ResNet34 provides robust feature extraction; only the decoder path needs to learn from the small camouflage dataset. This single decision was likely the largest factor in achieving a usable IoU.
Three-group augmentation pipeline — Groups target distinct perceptual variation axes: tone/color (CLAHE, brightness/contrast, gamma, HSV), blur type (Gaussian, median, motion), and signal degradation (noise, JPEG compression). Each group fires independently, producing different combinations each epoch without over-stacking transforms of the same perceptual category.
Color-coded mask extraction — Storing four semantic classes as distinct BGR color tuples in a single PNG avoids per-class file I/O at the cost of a correctness constraint: mask images must contain exact pixel values (no JPEG artifacts). The Dataset extracts binary channels via exact-color inRange matching at load time; the class_values parameter allows requesting class subsets without changing mask files.