This paper proposes a novel semantic segmentation algorithm based on learning a deconvolution network. The deconvolution network is built upon the convolutional layers of the VGG 16-layer network. It consists of deconvolution and unpooling layers that identify pixel-wise class labels and predict segmentation masks. The trained network is applied to each object proposal in an input image, and the final segmentation map is constructed by combining the results from all proposals. The proposed algorithm addresses the limitations of existing fully convolutional network-based methods by integrating a deep deconvolution network and proposal-wise prediction. It effectively handles objects of various scales and identifies detailed structures. The deconvolution network is trained only on the PASCAL VOC 2012 dataset and achieves the best accuracy (72.5%) among methods trained without external data through ensemble with the fully convolutional network. The network's architecture includes a convolution network for feature extraction and a deconvolution network for generating segmentation masks. The deconvolution network is a mirrored version of the convolution network, with multiple layers of unpooling, deconvolution, and rectification. The deconvolution process reconstructs the input image from its feature representation, enabling the network to capture detailed object structures. The deconvolution network is trained using a two-stage approach. In the first stage, object instances are cropped and used to train the network with fewer examples. In the second stage, object proposals are used to create more challenging examples. The network is trained using batch normalization to reduce internal covariate shift and improve optimization. The trained network is then used for inference, where it generates object proposals and combines their outputs to produce a final segmentation map. The results are further improved by ensembling with the FCN-8s network. The proposed method outperforms existing methods on the PASCAL VOC 2012 dataset, achieving the best accuracy among methods trained only on PASCAL VOC data. The deconvolution network's ability to capture detailed object structures and handle multi-scale objects makes it effective for semantic segmentation. The method's complementary characteristics with FCN-based approaches allow for better performance through ensemble. The network's performance is validated through experiments on the PASCAL VOC 2012 dataset, demonstrating its effectiveness in semantic segmentation.This paper proposes a novel semantic segmentation algorithm based on learning a deconvolution network. The deconvolution network is built upon the convolutional layers of the VGG 16-layer network. It consists of deconvolution and unpooling layers that identify pixel-wise class labels and predict segmentation masks. The trained network is applied to each object proposal in an input image, and the final segmentation map is constructed by combining the results from all proposals. The proposed algorithm addresses the limitations of existing fully convolutional network-based methods by integrating a deep deconvolution network and proposal-wise prediction. It effectively handles objects of various scales and identifies detailed structures. The deconvolution network is trained only on the PASCAL VOC 2012 dataset and achieves the best accuracy (72.5%) among methods trained without external data through ensemble with the fully convolutional network. The network's architecture includes a convolution network for feature extraction and a deconvolution network for generating segmentation masks. The deconvolution network is a mirrored version of the convolution network, with multiple layers of unpooling, deconvolution, and rectification. The deconvolution process reconstructs the input image from its feature representation, enabling the network to capture detailed object structures. The deconvolution network is trained using a two-stage approach. In the first stage, object instances are cropped and used to train the network with fewer examples. In the second stage, object proposals are used to create more challenging examples. The network is trained using batch normalization to reduce internal covariate shift and improve optimization. The trained network is then used for inference, where it generates object proposals and combines their outputs to produce a final segmentation map. The results are further improved by ensembling with the FCN-8s network. The proposed method outperforms existing methods on the PASCAL VOC 2012 dataset, achieving the best accuracy among methods trained only on PASCAL VOC data. The deconvolution network's ability to capture detailed object structures and handle multi-scale objects makes it effective for semantic segmentation. The method's complementary characteristics with FCN-based approaches allow for better performance through ensemble. The network's performance is validated through experiments on the PASCAL VOC 2012 dataset, demonstrating its effectiveness in semantic segmentation.