🍃 Leafers

A final year project of using deep learning CNN models for predicting Hong Kong plants images.

Leafers website | Google Play | GitHub

📊 Dataset

🌸 flower400

Dataset that contains 400 types of flowers occur in Hong Kong. (86880 images)
https://www.kaggle.com/r48n34/flower400

🌳 flower258

Dataset that contains 258 types of flowers occur in Hong Kong, group by genus. (86880 images)
https://www.kaggle.com/r48n34/flower258

🏞 flower yolov5

Using flower400 images, selecting few of representative for labeling with boxes.
https://www.kaggle.com/r48n34/flowers-yolov5

💪 Training params

📸 Images classifications

👠 Model choosing:
Tensorflow 2.8 EfficientNetV2 & EfficientNet & MobienetV3 series

🔪 Data spliting:
80% Training & 20% Validations

📈 Max scale mtehod:
Scale up all classes that below a number of the max images classes

📈 Max scale Flower400 / Flower258:
Training - 276550 images
Validations - 7369 images

🔍 Object detections

👠 Model choose:
yolov5 small

📈 Data:
Training - 1067 images
Validations - 100 images

🌲 Environments

Specs \ Env	Colab ENV	Local ENV
CPU	Intel(R) Xeon(R) CPU @ 2.30GHz	AMD Ryzen 5 5600X 6-core 3.7GHz (4.5 OC)
GPU	NVIDIA Tesla P100 16GB	MSI GeForce RTX 3060 Ti 12GB
RAM	24 GB	16 GB DDR4 3200
CUDA	11.2	11.5
TF version	2.7	2.8

💪 Training result

🌳 flower258

Code	Aug	Method	Scale	Train Top1	diff	Predict Top1	Predict Top5
10	max scale + RandAug	Official effNetv2b1	240, [0,255]	8436	-96	8340	9495
11	max scale + RandAug	MobienetV3	224, [0,255]	8073	-173	7900	9339
ex		9,10,11 Sum output			N/A	7963	9383
ex		9,10,11 Voting			N/A	7835	N/A
12	max scale + Nornal Aug	effNetv1b3	300, [-1,1]	8403	-637	7766	9273
14	max scale + RandAug	Official effNetvb2	260, [0,255]	8313	-175	8138	9442
15	max scale + RandAug	Official effNetvb0	224, [0,255]	8752	-90	8662	9583

🌸 flower400

Code	Aug	Method	Scale	Train Top1	diff	Predict Top1	Predict Top5
13	max scale + RandAug	Official effNetv2b3	300, [0,255]	7323	-100	7223	9279
16	max scale + RandAug	Official effNetv2b0	224, [0,255]	7316	-88	7228	9257
17	max scale + RandAug	13,16 stacking	300, [0,255]	7602	-51	7551	9375
18	max scale + CutMix	Official effNetv2b1	240, [0,255]	7363	-354	7009	9183
19	none	Official effNetb0	224, [0,255]	7417	-328	7089	8604
20	RandAug cont 19 model	Official effNetb0	224, [0,255]	7445	-390	7055	8482
21	max scale + RandAug	Official effNetv2b1	240, [0,255]	7417	-111	7306	9330
22	max scale + RandAug	13,16,21 stacking	300, [0,255]	7726	+5	7731	9475

🏞 YoloV5 Best.pt

train/box_loss	train/obj_loss	precision	recall	mAP_0.5	mAP_0.5:0.95	val/box_loss	val/obj_loss
0.023474	0.023782	0.9516	0.77108	0.87833	0.63541	0.031121	0.012252

💻 Applications development

🌐 Website :
React with Vite (Typescript, Firebase, tfjs, Mantine, redux RTK...)

🤖 Apps (Android) :
React native with expo (tfjs)

💻 Serverless API deploy :
Azure functions

📦 Frontend deploy:
(Old) Azure static web app => (New) Vercel

👍 Special Thanks

📷 Media:

• Startupbeat

📊 Data References:

• Hong Kong Herbarium
• HKCWW
• Produce Green HK

📑 Protocol References:

• Smart Flower Atlas
• Pl@ntNet