Multi-scale Interactive Network for Salient Object Detection

Abstract: Deep-learning based salient object detection methods achieve great progress. However, the variable scale and unknown category of salient objects are great challenges all the time. These are closely related to the utilization of multi-level and multi-scale features. In this paper, we propose the aggregate interaction modules to integrate the features from adjacent levels, in which less noise is introduced because of only using small up-/down-sampling rates. To obtain more efficient multi-scale features from the integrated features, the self-interaction modules are embedded in each decoder unit. Besides, the class imbalance issue caused by the scale variation weakens the effect of the binary cross entropy loss and results in the spatial inconsistency of the predictions. Therefore, we exploit the consistency-enhanced loss to highlight the fore-/back-ground difference and preserve the intra-class consistency. Experimental results on five benchmark datasets demonstrate that the proposed method without any post-processing performs favorably against 23 state-of-the-art approaches. The source code will be publicly available at https://github.com/lartpang/MINet.

• The paper introduces a multi-scale architecture that combines aggregate interaction modules to integrate adjacent feature levels for improved salient object detection.
• The model uses self-interaction modules to extract detailed multi-scale features from single convolutional blocks, enhancing feature preservation.
• The approach employs a novel consistency-enhanced loss to address class imbalance and consistently outperforms 23 state-of-the-art SOD models across five datasets.

Multi-scale Interactive Network for Salient Object Detection

The paper "Multi-scale Interactive Network for Salient Object Detection" presents a novel approach aimed at enhancing the efficacy of salient object detection (SOD) by addressing challenges related to variable scales and category ambiguities in salient objects. The proposed solution leverages multi-level and multi-scale features through innovative network architectures, namely aggregate interaction modules (AIMs) and self-interaction modules (SIMs).

Methodological Overview

The methodology introduces AIMs to effectively integrate features from adjacent levels, minimizing noise through low up-/down-sampling rates. This integration allows the model to more robustly manage various object scales by preserving more detailed information from shallow features. In parallel, SIMs in each decoder unit enable the extraction of multi-scale information from single convolutional blocks, enhancing the network's flexibility and detail preservation.

A significant component of addressing class imbalance and spatial prediction inconsistency is the novel consistency-enhanced loss (CEL) function. CEL emphasizes foreground-background differentiation and maintains intra-class consistency, offering an alternative to the traditional binary cross entropy loss which struggles with scale-induced class imbalances.

Empirical Results

The approach is benchmarked against 23 state-of-the-art SOD models across five datasets, demonstrating superior performance without post-processing. Strong numerical results highlight its competitive advantages: for instance, in terms of Maximum F-measure and Mean Absolute Error (MAE), the method consistently outperforms existing models, pointing to its efficacy in handling complex scale variations.

Implications and Future Directions

The proposed methodology offers practical benefits in fields requiring precise object boundary detection, such as image retrieval and non-photorealistic rendering. Theoretical contributions include improved integration strategies within neural networks that could be translated to other convolutional tasks needing multi-scale awareness.

Future research could extend the current architecture by exploring adaptive mechanisms within AIMs and SIMs to further improve computational efficiency and accuracy. Additionally, the exploration of CEL in other domains could yield insights into managing class imbalance and coherence optimally across a range of AI tasks.

In sum, the paper provides a well-founded contribution to the field of salient object detection, introducing innovative solutions and demonstrating their effectiveness through rigorous quantitative evaluation. The implications for both practical applications and theoretical advancements in AI are noteworthy, paving the way for further exploration and refinement.