Joint Task of Image Segmentation and Classification for Object Detection

Object detection is the task of localizing and classifying objects in a given image that are in a closed set of objects of interest. Deep neural networks have improved the performance of object detection in recent times. However, we observe that the networks often comprise many layers, large number of network parameters and require large training dataset. Based on these observations, we cast object detection as a joint task of image segmentation and classification to increase detection accuracy while reducing computational cost. The joint task is underpinned by the use of attention gates to highlight the multi-scale salient regions in an image. Subsequently, a component shrinking algorithm generates bounding box from object boundaries that results from the image segmentation. Efficiency of classification is gained by integrating the Inception and Xception modules in the proposed network. Extensive experiments were designed to validate the hypothesis and bench-mark the proposed network against state-of-the-art deep networks. The results indicate that the proposed network exhibits better segmentation and detection accuracy while using fewer number of network parameters.