Receptive Field বোঝা

Effective RF — empirical phenomenon, theoretical analysis-এর সাথে discrepancy। Object detection এর জন্য critical।

Theoretical vs effective:

• Theoretical — সব input pixel equal influence assume।
• Effective — gradient-flow weighted contribution।
• Center pixel — multiple paths, edges — fewer।
• Gaussian-shaped attention।

Why Gaussian shape?

• Random weight initialization।
• Center পাথ — convolution-এ direct contribution।
• Edge — fewer convolutional paths।
• Central limit theorem-এর effect।

Mathematical intuition:

• $N$-layer stack → effective RF ~ $\sqrt{N} \cdot \sigma$।
• Linear layer growth — sub-linear ERF।
• "Most learning happens in central region"।

Implications for detection:

• Small object → small RF needed → shallow features।
• Large object → big RF → deep features।
• Multi-scale architecture necessary।

FPN (Feature Pyramid Network):

• Lin et al. (২০১৭) — Detection-এ multi-scale।
• Top-down pathway — high-level feature low-resolution।
• Lateral connection — fine resolution merge।
• Each pyramid level — different effective RF।

FPN architecture:

P5 (deep, low-res, large RF)
  ↓ upsample
P4 (mid, mid-res, mid RF) ← lateral
  ↓ upsample
P3 (shallow, high-res, small RF) ← lateral
  ↓ upsample
P2 (shallowest, highest res)

Detection at each scale:

• P2/P3 — small object।
• P4/P5 — large object।
• Per-scale anchor optimal।

Modern advances:

• BiFPN (EfficientDet): bidirectional fusion।
• NAS-FPN: architecture search।
• PANet: bottom-up augmentation।

Receptive field design rules:

• Object size ≤ effective RF।
• Multi-scale features for varied object sizes।
• Atrous/dilation for compact RF expansion।
• Attention for adaptive RF।

Bangladesh use cases:

• License plate detection — wide aspect, multiple scales।
• Crop disease detection — leaf spots small।
• Crowd counting — distant heads small।
• Multi-scale FPN আদর্শ।

Visualization tools:

• Activation maximization — RF visualize।
• Gradient-based — effective RF measure।
• CAM — class-relevant region।

মূল উপলব্ধি: Theoretical RF — paper number। Effective RF — practical reality। Gap exists, design must account। FPN — multi-scale solution। Detection-এ universal। Bangladesh detection task — FPN backbone + task-specific augmentation। Effective RF awareness — better architecture decision।

Semantic segmentation — পুরো image-এ per-pixel class prediction। Resolution এবং context — দু'টোই দরকার।

Standard CNN problem:

• Pool layer — RF বাড়ায় কিন্তু resolution কমায়।
• $224 \times 224$ → $7 \times 7$ feature map।
• Per-pixel prediction এর জন্য — কীভাবে $7 \times 7$ থেকে $224 \times 224$ output?
• Upsampling spatial detail recover-এ কঠিন।

FCN approach (২০১৫):

• Long-Shelhamer-Darrell — Fully Convolutional Network।
• Pool-based downsample + transposed conv upsample।
• Skip connections — fine details preserve।
• Initial breakthrough কিন্তু blurry boundary।

DeepLab (২০১৪+):

• Chen et al. — Atrous (dilated) convolution।
• Pool-এর জায়গায় dilated conv।
• Resolution preserve + RF বাড়ে।
• Segmentation accuracy major jump।

Atrous Spatial Pyramid Pooling (ASPP):

• Multiple dilation rate parallel — 1, 6, 12, 18।
• Different scale context capture।
• Concatenate → fused feature।
• Multi-scale segmentation।

Implementation:

class ASPP(nn.Module):
    def __init__(self, in_ch, out_ch=256):
        super().__init__()
        self.branches = nn.ModuleList([
            nn.Conv2d(in_ch, out_ch, 1),
            nn.Conv2d(in_ch, out_ch, 3, padding=6, dilation=6),
            nn.Conv2d(in_ch, out_ch, 3, padding=12, dilation=12),
            nn.Conv2d(in_ch, out_ch, 3, padding=18, dilation=18),
        ])
        self.fuse = nn.Conv2d(out_ch * 4, out_ch, 1)
    def forward(self, x):
        feats = [b(x) for b in self.branches]
        return self.fuse(torch.cat(feats, 1))

Why dilation works:

• RF expand without resolution loss।
• Parameter cost minimal।
• Multi-scale natural।
• Boundary preservation better।

Limitations of dilation:

• Gridding artifact — sparse sampling।
• Memory cost — feature map বড় থাকে।
• Computation cost — high resolution-এ দীর্ঘ।
• Hybrid Atrous (HDC) — gridding mitigate।

Modern alternatives:

• U-Net: encoder-decoder + skip — medical imaging gold।
• SegFormer: Vision Transformer + lightweight head।
• Mask2Former: attention-based universal segmentation।

U-Net specifics:

• Encoder downsample — context।
• Decoder upsample — resolution।
• Skip connections — fine details।
• Medical imaging-এ standard।

Bangladesh segmentation use:

• Medical: tumor segmentation, organ delineation।
• Agriculture: crop boundary, disease area।
• Satellite: land use, building detection।
• Document: Bangla text region detection।

Evaluation metrics:

• IoU (Intersection over Union)।
• Dice coefficient।
• Per-class accuracy (class imbalance)।
• Boundary F-measure।

মূল উপলব্ধি: Segmentation-এ resolution + context dual demand। Pool resolution হত্যা করে — dilation alternative। DeepLab series বিপ্লব — atrous + ASPP। U-Net medical-এ gold। Modern transformer-based architecture rising। Bangladesh segmentation — domain-specific architecture choice + augmentation key।

ViT (Dosovitskiy ২০২০) — pure transformer image classification। CNN-এর সাথে fundamental architectural difference।

CNN RF growth — slow:

• Each conv layer linear RF increase।
• 50-layer ResNet — RF ~500 (theoretical), effective ~100।
• Hierarchical — local → global gradual।

ViT RF — instant global:

• Patch attention — every patch sees every patch।
• Layer 1 থেকেই full image RF।
• No locality bias।

Architecture:

• Image → 16×16 patches।
• $224 \times 224 \to 14 \times 14 = 196$ patches।
• Each patch → embedding vector।
• Transformer block stack।

Why "revolutionary":

• Single architecture for vision + language।
• Scale-friendly — bigger model, more data, better।
• Inductive bias minimal — flexibility।
• SOTA results on multiple benchmarks।

Trade-off — data hunger:

• CNN — image-friendly inductive bias।
• ViT — bias none, must learn from data।
• JFT-300M (Google internal) — ViT shines।
• ImageNet-1M — CNN comparable।

Empirical observations:

• Small data — CNN better।
• Large data — ViT scales better।
• Hybrid — CNN stem + ViT body — সবসময় ভাল।

Compute trade-off:

• Attention quadratic in patches।
• Long-range powerful কিন্তু expensive।
• Memory significant।

Modern ViT improvements:

• Swin Transformer: local windows + shifted।
• DeiT: distillation efficient training।
• BeIT: BERT-style pre-training।
• DINO: self-supervised।

Hybrid architectures:

• ConvNeXt — modernized CNN matching ViT।
• CoAtNet — Conv + Attention।
• EfficientNet ViT hybrid।

Modern best practice:

• ImageNet-1M pretrain — Swin/ConvNeXt comparable।
• Self-supervised (DINO) — ViT excellent।
• Multi-modal (CLIP) — ViT preferred।

Bangladesh deployment:

• Limited data — CNN preferred।
• Pre-trained ViT fine-tune — works।
• Edge device — CNN smaller।
• Cloud inference — ViT acceptable।

Future direction:

• Multi-modal foundation models।
• Vision + language unified।
• Few-shot learning।
• Generative + discriminative।

মূল উপলব্ধি: ViT — global RF instant from layer 1। CNN-এর hierarchical গঠন বদলে — direct attention। Data hungry কিন্তু scale powerful। Modern computer vision dual ecosystem — CNN, ViT, hybrid। Bangladesh-এ — task + data-এর উপর depend। Foundation model era — multi-modal, scaling, self-supervision।

Small object detection — computer vision-এর কঠিনতম problem। Drone agriculture-এ practical importance।

Problem analysis:

• Drone image: 4K (3840×2160) typical।
• Pest: 10-20 pixel (rice plant brown spot, leaf hopper)।
• Aspect: typically square।
• Density: clustered or sparse।

RF requirement:

• Pest size 20px — RF ≥ 30-50px (some context)।
• Standard ResNet — RF ~480 — overkill, miss small।
• Need: shallow + multi-scale।

Architecture choice:

• FPN backbone: ResNet50 + FPN।
• P2 (1/4) — small pest detection।
• P3-P5 — context for verification।
• RetinaNet style detector।

Patch-based training:

• 4K image → 512×512 patches।
• Overlap 64 pixel — boundary objects।
• Train on patch — efficient।
• Inference — sliding window।

Anchor design:

• Small anchor: 8, 16, 32 pixel।
• Aspect ratio: 1:1, 1:2, 2:1।
• Per-level anchor — small for P2, large for P5।

Loss specifics:

• Focal loss — class imbalance (background 99%)।
• GIoU — better small object localization।
• Hard example mining।

Augmentation:

transforms = A.Compose([
    A.RandomScale(scale_limit=0.3, p=0.5),
    A.RandomRotate90(p=0.5),
    A.HorizontalFlip(p=0.5),
    A.Mosaic(p=0.5),  # combine 4 images
    A.MixUp(p=0.3),
    A.RandomBrightnessContrast(0.2, 0.2),
    A.HueSaturationValue(20, 30, 20),
    A.GaussianBlur(blur_limit=3, p=0.3),
])

Mosaic augmentation:

• YOLOv4 introduced।
• 4 images combine — diverse context।
• Small object effective training।
• Critical for small pest।

Data labeling:

• Expert agronomist labeling।
• Bounding box tight।
• Multi-class: brown plant hopper, stem borer, etc.।
• Iterative active learning।

Bangladesh-specific challenges:

• Monsoon — drone image quality vary।
• Crop variety — Boro, Aman, Aush different appearance।
• Pest seasonality — annual pattern।
• Limited labeled data।

Pre-training strategy:

• ImageNet pre-trained backbone।
• Crop disease open dataset (PlantVillage)।
• Self-supervised on unlabeled drone images।
• Fine-tune on labeled subset।

Inference deployment:

• Drone onboard — Jetson Nano/Xavier।
• Quantized YOLOv5 small।
• Real-time 30 FPS।
• Cloud sync — flight data।

Field considerations:

• Drone altitude consistent — 5-10m।
• GPS coordinate per detection।
• Spray application targeted।
• Farmer mobile app — Bangla UI।

Validation strategy:

• Field-wise hold-out।
• Season-wise split।
• Variety-wise generalization test।
• Expert validation।

Economic value:

• Early detection — yield save।
• Targeted spray — pesticide reduce।
• Cost saving farmer।
• Environmental benefit।

মূল উপলব্ধি: Small object detection — RF careful balance। Multi-scale FPN + small anchors essential। Mosaic augmentation small object boost। Bangladesh agriculture — drone + AI promising। Practical pipeline — patch training + onboard inference + Bangla farmer interface। Receptive field theory practice transformation।