ConvBAMnet: An Improved Approach for Camouflage and Occlusion Detection

Abstract

Nowadays camouflage object detection is more challenging task as background and foreground objects are almost similar in color. Most of the time when the objects are fully or partially occluded along with camouflage then the problem is worst giving incorrect predictions. To address these issues, this paper proposes an algorithm ConvBAMnet which improves the feature extraction by introducing The Bottleneck Attention Module (BAM) layer. Convolutional layers, dense layers with dropout, bounding box regression, and attention processes work together to create a model that excels at identifying objects in difficult situations like obscured and camouflaged surroundings. BAM layers are specifically made to draw attention to specific channels and geographical locations within feature maps. The dataset considered over here is Moving camouflaged animal (MoCA) dataset. Two classes are considered for training purpose. Flat fish is highly camouflaged and Nile monitor is camouflaged as well as partially occluded. The problem of false and missed detection is improved giving better detection results by proposed method. The proposed model ConvBAMnet has been compared with ResNet 50 (Residual Neural Network 50) and VGG 16 (Visual Geometry Group 16) which were transfer learned on given dataset. The experimentation results indicate that proposed model has achieved an increase in accuracy compared to SOTA ResNet 50 and VGG 16.