利用細節圖和混合損失函數的顯著性偵測

Abstract

顯著性物件偵測（Salient Object Detection，SOD）在各種電腦視覺應用中扮演著至關重要的前處理步驟，包括視覺追蹤、影像描述、影像分割和影像辨識。本研究旨在通過結合各種技術來提高SOD的準確性。相較於僅使用顯著圖進行監督，我們提出一種不同的角度，以改進交互式雙流解碼器。這包括生成主體圖和細節圖，提供豐富的顯著訊息，來獲得精確的預測。使用者可以微調參數對比來進行客製化。此外，我們引入了一個對應的損失函數，可以分成兩個部分：像素級損失和物件級損失。通過使用這種雙級損失函數，我們可以訓練出更準確的模型，以獲得更精確的預測。最後，我們會在3個資料集上與其他11個目前最流行的SOD演算法去做比較，利用3個既有的指標進行評估。我們的方法都是最佳或是前幾名的成績

Salient Object Detection (SOD) represents a crucial preprocessing step in various computer vision applications, including visual tracking, image captioning, image segmentation, and image recognition. This research aims to enhance the accuracy of SOD by incorporating innovative techniques. Instead of relying solely on saliency maps for supervision, we propose an advanced approach to improve the interactive two-stream decoder. This involves the strategic generation of body maps and detail maps, providing substantial salient information for precise predictions. Users can fine-tune parameter contrasts for customization. Additionally, we introduce a corresponding loss function which can be divided into two part: pixel-level loss and object-level loss. By employing this two-level loss function, we can train a more accurate model to obtain more precise predictions. Compared to state-of-the-art SOD algorithms, our method consistently outperforms competitors across three datasets evaluated with three established metrics.