通過深度學(xué)習分析動物的面部表情是本研究的主要內(nèi)容。由于羊面部表情的人工評價缺乏準確性,耗時且單調(diào)。因此面部表情的疼痛水平估計是綿羊生命的有效和可靠的標記。一方面深度學(xué)習的基礎(chǔ)是計算機視覺中卷積神經(jīng)網(wǎng)絡(luò)的最新進展,有助于快速準確地對面部表情進行分類。但另一方面,由于輸入中的高斯和脈沖噪聲,卷積神經(jīng)網(wǎng)絡(luò)易受小樣本的影響。在本方法第一階段中,首先要消除數(shù)字圖像處理中的組合高斯和脈沖噪聲。由于保留圖像細節(jié)和抑制噪聲是具有挑戰(zhàn)性的問題,為此,一種結(jié)合卷積神經(jīng)網(wǎng)絡(luò)新型中值濾波器濾波器被用于處理高斯和椒鹽噪聲。以前的方法是依賴程序,一些用于處理脈沖噪聲,另一些用于處理高斯噪聲。首先消除高斯和脈沖噪聲的是通過采用3×3和5×5窗口大小的中值濾波器來檢測具有噪聲抑制的脈沖噪聲。隨后在第二步中,通過殘差學(xué)習去噪卷積神經(jīng)網(wǎng)絡(luò)去除高斯噪聲。在數(shù)字圖像處理領(lǐng)域,學(xué)習和去噪性能非常必要。去噪卷積神經(jīng)網(wǎng)絡(luò)還具有有效處理具有未知的噪聲水平的高斯噪聲。在第二階段時,通過監(jiān)測綿羊的生活實現(xiàn)對自然習性的充分評估,這對于管理至關(guān)重要。在這項研究中,我們提出了一種羊臉數(shù)據(jù)集的框架,它使用轉(zhuǎn)移學(xué)習和微調(diào)來自動分類正常和異常的羊臉...

【文章頁數(shù)】：79 頁

【學(xué)位級別】：碩士

【文章目錄】：
ABSTRACT
摘要
INTRODUCTION
    1.1.RESEARCH BACKGROUND
    1.2.LITERATURE SURVEY
        1.2.1.Overview
    1.3.LAYOUT OF THESIS
FUNDAMENTALS OF DEEP LEARNING
    2.1.SCOPE
    2.2.CONVOLUTIONAL NEURAL NETWORK
        2.2.1.Framework
        2.2.2.Convolution Layer
        2.2.3.Activation functions
        2.2.4.Pooling Layer
        2.2.5.Fully Connected Layer
        2.2.6.Logistic Regression
        2.2.7.Optimization Algorithms
        2.2.8.Learning Rate Decay
        2.2.9.Tuning Process
        2.2.10.Batch Normalization
        2.2.11.Softmax Regression
        2.2.12.Deep Learning Projects Strategy
        2.2.13.Transfer Learning
    2.3.NOISE REMOVAL ALGORITHMS
        2.3.1.Median Filter
        2.3.2.Adaptive Center Weighted Median Filter
        2.3.3.Decision-Based Algorithm
        2.3.4.Noise Adaptive Fuzzy Switching Median Filter
        2.3.5.Efficient Edge-Preserving Algorithm for Removal of Salt-and-Pepper Noise
        2.3.6.De-noising Convolutional Neural Network
    2.4.DISCUSSION
METHODOLOGY
    3.1.SCOPE
    3.2.GAUSSIAN AND IMPULSE DENOISING ALGORITHM
        3.2.1.Proposed Method Overview
        3.2.2.Step 1
        3.2.3.Step 2
    3.3.AUTOMATED SHEEP FACIAL EXPRESSION PIPELINE
        3.3.1.Sheep face dataset
        3.3.2.Transfer Learning
        3.3.3.Custom read function
        3.3.4.Data Augmentation
        3.3.5.Regularization
        3.3.6.Stochastic Gradient Descent with Momentum(SGDM)
        3.3.7.Fine-Tuning
    3.4.PAIN RATING SCALES
    3.5.DISCUSSION
EXPERIMENTAL RESULTS AND DESCUSSION
    4.1.SCOPE
    4.2.NOISE REMOVAL ALGORITHM SIMULATION AND RESULTS
    4.3.SHEEP FACIAL EXPRESSION EXPERIMENTAL RESULTS
        4.3.1.Tools and Setup
        4.3.2.Normal and Abnormal Sheep Face Results
    4.4.SHEEP PAIN RATING SCALES RESULTS
    4.5.DISCUSSION
CONCLUSIONS AND FUTURE WORK
REFERENCES
PAPERS
Acknowledgement



本文編號：3789539