We propose a novel Quantum Convolutional Neural Network (QCNN) architecture that integrates Residual Encoding Blocks (R \(^{2}\) ) and Diagonal Residual Layers within an alternating feature encoding framework. Our design adapts classical convolution and pooling principles to quantum circuits by employing parameterized two-qubit operations, mid-circuit measurements for quantum pooling, and ancilla-driven skip connections to mitigate through barren plateaus and unexplored information within the system. The inclusion of diagonal residual layers facilitates entanglement between non-adjacent qubits, enhancing global feature extraction and information propagation in deeper circuits. We evaluated our model on a 3 \(\times \) 3 Tic-Tac-Toe classification task, where each board configuration is encoded into a 9-qubit i.e. quantum bit, quantum state and the model predicts the final winner (White or Black). Experimental results demonstrate that the proposed QCNN achieves high classification accuracy while significantly reducing parameter count and circuit depth compared to baseline quantum classifiers, highlighting potential of quantum-machine learning in the upcoming time. Check repository: https://github.com/f361015/Quantum-CNN-TicTacToe .