Modified Faster R-CNN with Applications to Cat and Dog Image Detection

Abstract

隨著深度學習技術的快速發展，神經網絡在物件檢測應用的範圍和性能上不斷改進，取得了顯著的進展。本論文基於 Faster R-CNN 框架，通過調整參數和卷積神經網絡，應用於 Kaggle 數據集中的貓狗圖像檢測。通過觀察性能變化並使用統計重採樣方法來確保數據集對模型精度和召回率的影響，論文展示了重採樣方法和參數調整如何影響模型的精度和召回率。在調整到最佳參數後，論文展示了基於 ResNet 的 Faster R-CNN 模型在物件特徵提取和邊界框回歸中的有效性，並比較了單階段物件辨識與兩階段物件辨識的精度差異。實驗結果表明，作為 Faster R-CNN 模型中特徵提取卷積神經網絡的 ResNet 在該數據集上表現出色，且兩階段物件辨識模型在此數據集上有較好的精度表現。

With the rapid development of deep learning technology, neural networks have continuously improved in both the scope and performance of object detection applications, achieving significant advancements. This thesis is basedon the Faster R-CNN framework, altering parameters and convolutional neural networks, and applies it to detecting cat and dog images in the Kaggle dataset. By observing performance changes and employing statistical resampling methods to ensure the precision and recall of the dataset's impact on the model, the thesis demonstrates how resampling methods and parameter adjustments affect model precision and recall. After adjusting for optimal parameters, the effectiveness of the ResNet-based Faster R-CNN model in object feature extraction and bounding box regression, and compares the accuracy differences between one-stage and two-stage object detection. Experimental results indicate that ResNet, used as the feature extraction convolutional neural network in the Faster R-CNN model, performs excellently on this dataset, and the two-stage object detection model exhibits better accuracy performance on this dataset.