Evaluation on 2,513 Unseen Test Samples

Model Performance & Analytics

Detailed quantitative evaluation, confusion matrix breakdown, precision/recall metrics, and training convergence logs for the GreenSort deep learning vision model.

Test Accuracy
92.04%
2,313 of 2,513 test items correct

Test Loss
0.2141
Binary crossentropy score

Total Dataset

25,077

22,564 train · 2,513 test

Training Duration

~101 min

15 epochs (~6.7 min / epoch)

🎯

Key Benchmark Finding: "The proposed CNN-based waste classification system achieved 92.04% accuracy on 2,513 previously unseen test images, with F1-scores of 0.93 for Organic waste and 0.91 for Recyclable waste."

Error Analysis

Confusion Matrix (Test Set)

N = 2,513

Actual Class

Predicted Class

Predicted O
(Organic)

Predicted R
(Recyclable)

Actual O
(Organic)

1,300 92.8% (True O)

101 7.2% (False R)

Actual R
(Recyclable)

99 8.9% (False O)

1,013 91.1% (True R)

Total Correct: 2,313 (92.04%)

Total Misclassified: 200 (7.96%)

Per-Class Metrics

Classification Report

Stratified Split

Class / Metric	Precision	Recall	F1-Score	Support
Organic (O)	0.93	0.93	0.93	1,401
Recyclable (R)	0.91	0.91	0.91	1,112
Accuracy	Overall Test Accuracy		0.92	2,513
Macro Avg	0.92	0.92	0.92	2,513
Weighted Avg	0.92	0.92	0.92	2,513

Organic (O) Insights: 93% of items classified as Organic were truly Organic, with only 101 misclassified into the recyclable stream.

Recyclable (R) Insights: 91% recall ensures minimal recyclable materials are lost to organic landfill streams.

Data Pipeline

Dataset Composition & Split Distribution

25,077 Total Images

Partition Breakdown (80/20 Train/Val Split)

72%

18%

10%

Train: 18,051 (72%)

Val: 4,513 (18%)

Test: 2,513 (10%)

Split Partition	Organic (O)	Recyclable (R)	Total Images	Ratio
TRAIN Set	12,565	9,999	22,564	89.98%
↳ Training Sub-split (80%)	~10,052	~7,999	18,051	71.98%
↳ Validation Sub-split (20%)	~2,513	~2,000	4,513	18.00%
TEST Set (Holdout)	1,401	1,112	2,513	10.02%
Full Dataset Total	13,966 (55.7%)	11,111 (44.3%)	25,077	100.0%

Class Balance Analysis

The dataset exhibits a natural, moderate class distribution (55.7% Organic vs 44.3% Recyclable) mirrored identically across both training and test sets.

🍃 Organic Waste (13,966 items) 55.68%

♻️ Recyclable Waste (11,111 items) 44.32%

📐 Tensor & Pipeline Dimensions

Batch Size: 32 samples per step
Tensor Shape: (32, 224, 224, 3) float32
Target Label: (32,) binary integer (0 = O, 1 = R)
Normalization: In-graph Rescaling(1/255)

Training Convergence

15-Epoch Training & Validation Progression

Adam (lr=0.001)

Epoch	Training Accuracy	Validation Accuracy	Training Loss	Validation Loss	Status
Epoch 8	88.28%	88.81%	0.2964	0.2918	Stable
Epoch 9	88.44%	88.65%	0.2892	0.2931	Learning
Epoch 10	88.65%	89.54%	0.2837	0.2748	Milestone
Epoch 11	89.18%	88.57%	0.2752	0.2983	Generalizing
Epoch 12	89.68%	88.37%	0.2631	0.2993	Refining
Epoch 13	89.68%	89.36%	0.2600	0.2727	Refining
Epoch 14	90.08%	90.01% ★	0.2541	0.2639 ★	Best Val Accuracy
Epoch 15	90.72%	89.59%	0.2411	0.2825	Final Epoch

Final Training Accuracy 90.72%

Peak Validation Accuracy 90.01%

Final Unseen Test Accuracy 92.04%