Lightweight single image deraining algorithm incorporating visual saliency

Abstract

Deep learning (DL) methods have achieved excellent performance in the task of single image rain removal, however, there are still some challenges, such as artifact remnant, background over-smooth, and more and more complex and heavy-weight network architecture. Due to too heavy-weight network to suit outdoor detection devices or mobile devices, therefore, we propose a light-weight single image deraining algorithm incorporating visual attention saliency mechanisms (LDVS). The network consists of dilation convolution module, the convolutional block attention module (CBAM), and gated recursive unit module. Specifically, rain steaks feature maps are extracted by the combinations of dilated convolution with CBAM, which also facilitates accurate localisation of rain steak location, and then the three gated recursive units is cascaded to remove steaks stage by stage. The dilated convolution module and CBAM are used to reduce network's weight size and retain the rain removal result, thus our LDVS method belongs to the lightweight with only 50703 parameters. Extensive experiments on synthetic and real-world datasets have demonstrated that our method outperforms the baseline both under qualitative and quantitative analysis. Under the same rain removal result, our method is less time cost and less burden.

Joint Semantic Deep Learning Algorithm for Object Detection under Foggy Road Conditions

“…Comprehensive experiments confirm that JSFR outperforms other methods for target detection on synthetic and real image in both foggy and normal conditions. In the future, we will further improve the accuracy of the algorithm and reduce network parameters to meet the needs of landing implementation [44]. Secondly, we will continue to deeply study the application scenarios such as video detection of bus traffic roads and vehicle or pedestrian tracking.…”

SemID: Blind Image Inpainting with Semantic Inconsistency Detection

“…Meanwhile, training on Paris StreetView and Places2 takes approximately 30 000 iterations in total (the first stage costs 10 000 iterations). We compare our method against some advanced conventional inpainting methods and state-of-the-art blind inpainting approaches, including CA [7] , GC [8] , HiFill [6] , VCNet [12] , and TransCNN-HAE [32] . For a fair comparison, CA, GC, and HiFill are all equipped with our IDN via sequential connections.…”

“…They first simulate multiple degradation patterns (e.g., graffiti and image stitching) and design a robust model that can identify degraded regions from semantic differences between contamination and surroundings before restoration. Zhao et al [32] proposed a one-stage hybrid autoencoder architecture, TransCNN-HAE, which exploits the powerful long-range context modeling capabilities of the transformer to avoid the possible degradation of inpainting performance from mask prediction errors. By contrast, the generalization capability of the model is considerably improved.…”

“…11c, replacing GC with vanilla convolutions. In the absence of MSCAM to [7] 0.9104 16.57 0.6210 27.09 0.9137 18.16 0.6387 78.65 0.7029 14.35 0.6281 33.82 GC [8] 0.8336 14.80 0.4161 22.52 0.7909 17.06 0.5349 66.17 0.7248 13.34 0.3497 31.85 HiFill [6] 0.9456 21.06 0.6076 28.60 0.9589 21.53 0.5843 63.36 0.9007 18.65 0.4900 11.24 VCNet [12] 0.9566 24.17 0.6691 16.88 0.9648 24.06 0.7272 55.93 0.9737 23.05 0.6854 14.65 TransCNN-HAE [32] − 3 to provide specific comparisons.…”

A Visual Target Representation using Saliency Detection Approach

“…Image enhancement is required to close the regions and remove some noises [19]. Therefore, noise removal [20], mathematical morphology operator [21], lightweight operators including dilation and closing [22], are applied to enhance the result of the edge detection process. Then filling function is applied, to fill the connected blobs in the enhanced image.…”