Convolutional Neural Network Design

Model Architecture & Engineering

An in-depth breakdown of layer dimensions, parameter counts, optimization strategies, and the parameter-efficient GlobalAveragePooling2D classifier design.

Trainable Parameters
421,441
1.61 MB active weights

Total Model Footprint
4.82 MB
1,264,325 total parameters

Input Resolution

224 × 224

3 RGB Color Channels

Feature Depth

256 Filters

4 Conv-Pool Stages

Key Optimization

GlobalAveragePooling2D vs. Traditional Flattening

Overfitting Prevention

❌ Traditional Flatten Approach

Flatten Layer (36,864 Features)

Flattening the final 12 × 12 × 256 feature map produces 36,864 linear inputs, requiring 4.72 Million parameters for the dense layer alone (over 5.1M total parameters).

Dense Layer Params: 4,718,720

High risk of severe overfitting on training textures
Massive memory and disk storage requirements
Sluggish inference times on edge/CPU environments

✅ GreenSort GlobalAveragePooling2D

Spatial Average Pooling (256 Features)

Computes the average activation across each feature map, directly condensing spatial dimensions to 256 values. The Dense layer requires only 32,896 parameters (~99.3% parameter reduction).

Dense Layer Params: 32,896 (99.3% reduction)

Significantly improved unseen generalization (92.04% Test Acc)
Eliminates spatial location bias in waste photos
Compact 4.82 MB file size enables rapid startup (<50ms)

Layer Specifications

Layer-by-Layer Network Topology

Keras Sequential Model

#	Layer Name	Layer Type	Output Shape	Kernel / Details	Param #
0	`input_layer`	Input / RGB	`(None, 224, 224, 3)`	Standard RGB Image	0
1	`rescaling_5`	Rescaling	`(None, 224, 224, 3)`	Scale pixels to [0, 1]	0
2	`conv2d_37`	Conv2D	`(None, 222, 222, 32)`	32 filters, 3×3, ReLU	896
3	`max_pooling2d_37`	MaxPooling2D	`(None, 111, 111, 32)`	Pool size 2×2, Stride 2	0
4	`conv2d_38`	Conv2D	`(None, 109, 109, 64)`	64 filters, 3×3, ReLU	18,496
5	`max_pooling2d_38`	MaxPooling2D	`(None, 54, 54, 64)`	Pool size 2×2, Stride 2	0
6	`conv2d_39`	Conv2D	`(None, 52, 52, 128)`	128 filters, 3×3, ReLU	73,856
7	`max_pooling2d_39`	MaxPooling2D	`(None, 26, 26, 128)`	Pool size 2×2, Stride 2	0
8	`conv2d_40`	Conv2D	`(None, 24, 24, 256)`	256 filters, 3×3, ReLU	295,168
9	`max_pooling2d_40`	MaxPooling2D	`(None, 12, 12, 256)`	Pool size 2×2, Stride 2	0
10	`global_avg_pool`	GlobalAveragePooling2D	`(None, 256)`	Spatial Average Pooling	0
11	`dense_22`	Dense	`(None, 128)`	128 Units, ReLU	32,896
12	`dropout_11`	Dropout	`(None, 128)`	Drop probability = 0.5	0
13	`dense_23`	Dense (Output)	`(None, 1)`	Sigmoid Binary Activation	129
Total Trainable Parameters					421,441 (1.61 MB)
Optimizer State (Adam Moments)					842,884 (3.22 MB)
Grand Total Parameters					1,264,325 (4.82 MB)

Hyperparameters

Training Configuration

Optimizer Adam (Adaptive Moment Estimation)

Learning Rate ($\alpha$) 0.001

Loss Function Binary Crossentropy

Batch Size 32 Images per Step

Training Epochs 15 Total Epochs

Steps per Epoch 565 Batches

Total Training Time ~101 minutes (~6.7 min / epoch)

Evaluation Latency 79 Test Steps in ~23 seconds

Production Pipeline

Inference Preprocessing & Decision Logic

Image Decoding & RGB Conversion:
Accepts JPEG, PNG, WEBP, BMP up to 15 MB and converts to 3-channel RGB.

High-Quality Resizing:
Resized to (224, 224) using Lanczos interpolation preserving edges and textures.

Batch Vectorization & Scaling:
Expanded to (1, 224, 224, 3). The in-graph Rescaling(1/255) normalizes pixel values without external scaling dependencies.

Sigmoid Decision Boundary:
Output $\sigma(z) \in [0, 1]$ represents recyclable probability:
• If $\sigma(z) \ge 0.5 \implies$ Class R (Recyclable), Confidence $= \sigma(z)$
• If $\sigma(z) < 0.5 \implies$ Class O (Organic), Confidence $= 1.0 - \sigma(z)$