# 2021/04/27 Agricultural project
###### tags: `Agricultural project`

## Leaf detection

### model

yolo_v2

### Data

Trainig: 221
testing: 139

### 目前實驗概況

目前loss最低約為1.97左右

||原圖|預測後圖片|
|--|--|--|
|1|||
|2|||
|3|||
|4|||
|5|||

上表為目前資料訓練後的預測結果
第一列為最理想之預測結果，第二列右下的box中出現重複的現象，第三與第四列則是將樹幹、草地等其他材質判斷為葉子，第五列為不佳的預測結果

### 問題與解決方法

1.重複框(上表2列) : 加入IOU判斷，若有重複框的狀況則刪除較小者
2.框到其他材質(上表3、4列) : 將樹幹、草地、其他雜草框做負樣本訓練
3.預測不佳(上表5列) : 增加訓練資料並修改訓練參數再做測試

---
## 網頁

撰寫網頁用的mask功能，能將輸入的圖片依照mask去背

|原圖|結果|
|--|--|
|||

---
## APP
### iOS
#### Step.1 初始頁面
#### Step.2 按下選取圖片選擇拍照或相簿照片
#### Stwp.3 拍攝照片(對準正中間)
#### Step.4 自動裁切並顯示辨識結果(固定位置無法調整)
#### Bouns：左上角選單頁面
---
### Android
#### Step 1. 初始畫面
#### Step 2. 選擇照片
#### Step 3. 裁切相片
#### Step 4. 上傳相片至伺服器
#### Step 5. 查看辨識結果
#### 用藥建議
#### 諮詢專線

---
## Plant Pathology 2021
## Dataset
Total 18632 iamges, 12 classes
All original labels in `train.csv` :
| Label | Count |
| ------------------------------- | ----- |
| scab | 4826 |
| healthy | 4626 |
| frog_eye_leaf_spot | 3181 |
| rust | 1860 |
| complex | 1602 |
| powdery_mildew | 1184 |
| scab frog_eye_leaf_spot | 686 |
| scab frog_eye_leaf_spot complex | 200 |
| frog_eye_leaf_spot complex | 165 |
| rust frog_eye_leaf_spot | 120 |
| rust complex | 97 |
| powdery_mildew complex | 87 |
### Labels
There are only 4 diseases and ``complex``, ``healthy`` in this dataset.
``complex`` means that the disease is too complicated to be recognized
> Unhealthy leaves with too many diseases to classify visually will have the complex class, and may also have a subset of the diseases identified.

| Label | Count |
| ------------------------------- | ----- |
| healthy | 4624 |
| scab | 5712 |
| frog_eye_leaf_spot | 4352 |
| complex | 2151 |
| rust | 2077 |
| powdery_mildew | 1271 |

## Experiment

training data: 14905
testing data:   3727
epoch:            80
batch_size:       32

#### 1. Single-label method + Single-label dataset
* Use Label Powerset to define labels.
* Contain 12 labels in full dataset.
* ACC : 0.8026
#### 2. Multi-label method + Single-label dataset
* Same labels rule as method 1
* Only change model into multi-label mode.
* ACC : 0.8133
#### 3. Multi-label method + Mingle-label dataset
* 6 labels (4 diseases + complex and healthy)
* scab frog_eye_leaf_spot complex is seen to 3 labels.
* ACC : 0.8484
---
## 硬體
與廠商討論需要的設備