#pragma region template
#include <bits/stdc++.h>

#include <atcoder/dsu>
#include <atcoder/maxflow>

using namespace std;

// clang-format off
using ll  = long long;    using vl   = vector<ll>;  using vvl   = vector<vl>;
using ld  = long double;  using vld  = vector<ld>;  using vvld  = vector<vld>;
using pll = pair<ll, ll>; using vpll = vector<pll>; using vvpll = vector<vpll>;
using vb = vector<bool>; using vvb = vector<vector<bool>>;
using vs = vector<string>;
using mll = map<ll, ll>;
template <class T> using V = vector<T>; template <class T> using VV = V<V<T>>; template <class T> using VVV = V<VV<T>>;
template <class T> using max_heap = priority_queue<T>;
template <class T> using min_heap = priority_queue<T, vector<T>, greater<T>>;
constexpr ll inf = 3001001000100100100LL;
#define endl '\n'

#define  _overload(_1, _2, _3, name, ...) name
#define   rep(...) _overload(__VA_ARGS__,   _rep,   _rep2,)(__VA_ARGS__)
#define  repc(...) _overload(__VA_ARGS__,  _repc,  _repc2,)(__VA_ARGS__)
#define  repr(...) _overload(__VA_ARGS__,  _repr,  _repr2,)(__VA_ARGS__)
#define reprc(...) _overload(__VA_ARGS__, _reprc, _reprc2,)(__VA_ARGS__)
#define   _rep(i,k,n) for(ll i=(k)  , i##_xxxx=(n); i < i##_xxxx; ++i)
#define  _repc(i,k,n) for(ll i=(k)  , i##_xxxx=(n); i <=i##_xxxx; ++i)
#define  _repr(i,k,n) for(ll i=(n)-1, i##_xxxx=(k); i >=i##_xxxx; --i)
#define _reprc(i,k,n) for(ll i=(n)  , i##_xxxx=(k); i >=i##_xxxx; --i)
#define   _rep2(i,n)   _rep(i,0,n)
#define  _repc2(i,n)  _repc(i,1,n)
#define  _repr2(i,n)  _repr(i,0,n)
#define _reprc2(i,n) _reprc(i,1,n)

#define rall(o) rbegin(o), rend(o)
#define  all(o)  begin(o),  end(o)
template <class C> ll sz(const C& c) { return static_cast<ll>(c.size()); }
template <class T> bool chmax(T& m, const T& v){ if (m < v){ m = v; return true; } return false; }
template <class T> bool chmin(T& m, const T& v){ if (v < m){ m = v; return true; } return false; }
template <class T> T cdiv(const T& a, const T& b){ return (a + b - 1) / b; }
template <class T> T rdiv(const T& a, const T& b){ return (a + b / 2) / b; }
template <class T, class S>             string join(const T& v, const S& sep                  ){ stringstream ss; bool f = false; for (const auto& e : v){ if (f) ss << sep; f = true; ss << e;} return ss.str(); }
template <class T, class S, class... U> string join(const T& v, const S& sep, const U& ...args){ stringstream ss; bool f = false; for (const auto& c : v){ if (f) ss << sep; f = true; ss << join(c, args...); } return ss.str(); }
ostream& operator<<(ostream& os, const __int128_t& val){ os << (ll)val; return os; }

// template <class Cap, class Edge = typename atcoder::mf_graph<Cap>::edge>
ostream& operator<<(ostream& os, const atcoder::mf_graph<__int128>::edge& e) {
    auto [from, to, cap, flow] = e;
    os << '(' << from << "->" << to << " : " << flow << "/" << cap << ')';
    return os;
}

template <class T>            ostream& operator<<(ostream& os, const vector<T>&            seq){ os << '[' << join(seq, ",") << ']'; return os; }
template <class T>            ostream& operator<<(ostream& os, const vector<vector<T>>&    seq){ os << '[' << join(seq, ",\n ") << ']'; return os; }
template <class T>            ostream& operator<<(ostream& os, const deque<T>&             seq){ os << '[' << join(seq, ",") << ']'; return os; }
template <class T>            ostream& operator<<(ostream& os, const set<T>&               seq){ os << '{' << join(seq, ",") << '}'; return os; }
template <class T , class TH> ostream& operator<<(ostream& os, const unordered_set<T, TH>& seq){ os << '{' << join(seq, ",") << '}'; return os; }
template <class TK, class TV> ostream& operator<<(ostream& os, const map<TK, TV>&          seq){ os << '{'; bool f = false; for (const auto& e : seq){ if (f) os << ','; f = true; os << e.first << ":" << e.second; } os << '}'; return os; }
template <class T1, class T2> ostream& operator<<(ostream& os, const pair<T1, T2>&          pa){ os << '(' << pa.first << ',' << pa.second << ')'; return os; }

#if LOCAL
#define _overload5(_1, _2, _3, _4, _5, name, ...) name
#define debug(...) _overload5(__VA_ARGS__, _d5, _d4, _d3, _d2, _d1, )(__VA_ARGS__)

#define _pp0(x, ...) #x " =", x
#define _pp(x, ...) ", "#x " =", x
//  ,pp(__VA_ARGS__)

#define _d1(x1) _debug(_pp0(x1), __LINE__)
#define _d2(x1, x2) _debug(_pp0(x1), _pp(x2), __LINE__)
#define _d3(x1, x2, x3) _debug(_pp0(x1), _pp(x2), _pp(x3), __LINE__)

// #define debug(...) _debug(__VA_ARGS__, __LINE__)
#else
#define debug(...)
#endif

void print() { std::cout << '\n'; }
template <class S>             void  print(const S& a){ std::cout << a << '\n'; }
template <class S>             void _debug(const S& a){ std::cerr << "(L:" << std::setw(3) << a << ")\n"; }
template <class S, class... T> void  print(const S& a, const T&... args){ std::cout << a << ' ';  print(args...); }
template <class S, class... T> void _debug(const S& a, const T&... args){ std::cerr << a << ' '; _debug(args...); }

struct setup_main { setup_main() { std::cin.tie(nullptr); std::ios::sync_with_stdio(false); std::cout << fixed << setprecision(15); } } setup_main_;
// clang-format on
#pragma endregion


template <class Cost>
struct PSP {
    struct Variable {
        ll id, flip;
        Variable(ll i = 0, bool f = false) : id(i), flip(f) {}
        Variable operator~() { return Variable(id, !flip); }
    };
    using S = Variable;
    ll n;
    vector<vector<Cost>> cost1;
    vector<vector<vector<vector<Cost>>>> cost2;
    vector<Variable> var;

    PSP(){};
    PSP(ll n_) : n(n_) {
        cost1.resize(n, vector<Cost>(2, 0));
        cost2.resize(n, vector(n, vector(2, vector<Cost>(2, 0))));
        var.reserve(n);
        rep(i, n) var.emplace_back(i);
    };

    void add_constraint(S x, Cost cost) { cost1[x.id][!x.flip] += cost; }
    void add_constraint(S x, S y, Cost cost) {
        if (x.id > y.id) swap(x, y), cost = -cost;
        cost2[x.id][y.id][!x.flip][!y.flip] += cost;
    };

    Cost solve() {
        Cost sum = 0;

        // find flips
        vector<vector<Cost>> cost2_rest(n, vector<Cost>(n));
        atcoder::dsu d(2 * n);
        rep(i, n) rep(j, n) {
            vector<vector<Cost>>& v(cost2[i][j]);
            Cost c = v[0][1] + v[1][0] - v[0][0] - v[1][1];
            cost1[i][0] += v[0][0];
            cost1[i][1] += v[1][0];
            if (c < 0) {
                cost1[j][1] += v[0][1] - v[0][0];
                cost2_rest[i][j] -= c;
            } else {
                cost1[j][1] += v[1][1] - v[1][0];
                cost2_rest[i][j] += c;
            }


            if (c > 0) { // 劣モジュラ
                d.merge(i, j);
                d.merge(i + n, j + n);
            } else if (c < 0) { // 非劣モジュラ
                d.merge(i, n + j);
                d.merge(n + i, j);
            } // c=0ならflipは自由
        }

        rep(i, n) if (d.same(i, n + i)) return -1;

        vl flip(2 * n, -1);
        rep(i, 2 * n) {
            ll p = d.leader(i);
            if (flip[p] == -1) flip[p] = 0;
            flip[i] = flip[p];
            flip[d.leader((n + i) % (2 * n))] = 1 - flip[p];
            debug(flip, i);
        }

        debug(sum);
        debug(flip);
        debug(cost1);
        debug(cost2_rest);

        // min-cut
        atcoder::mf_graph<Cost> graph(n + 2);
        ll s = n, t = n + 1;
        rep(i, n) { // cost1
            if (!flip[i]) swap(cost1[i][0], cost1[i][1]);
            if (cost1[i][0] >= cost1[i][1]) {
                graph.add_edge(s, i, cost1[i][0] - cost1[i][1]);
                sum += cost1[i][1];
            } else {
                graph.add_edge(i, t, cost1[i][1] - cost1[i][0]);
                sum += cost1[i][0];
            }
        }
        rep(i, n) rep(j, n) {
            if (flip[j])
                graph.add_edge(i, j, cost2_rest[i][j]);
            else
                graph.add_edge(i, j, -cost2_rest[i][j]);
        }

        ll f = graph.flow(s, t);

        debug(sum, f);
        debug(graph.edges());
        debug(graph.min_cut(s));
        return sum + f;
    }
};


template <class T>
struct Compress {
    vector<T> inv;
    Compress(vector<T> a = {}) : inv(a) {}

    void push(T val) { inv.push_back(val); }
    void compress() {
        sort(all(inv));
        inv.erase(unique(all(inv)), inv.end());
    }

    ll zip(T val) { return lower_bound(all(inv), val) - inv.begin(); }

    T unzip(ll idx) {
        assert(0 <= idx and idx < sz(inv));
        return inv[idx];
    }

    size_t size() { return inv.size(); }
};


#ifdef LOCAL
const ll M = 10;
#else
const ll M = 500000;
#endif

void solve(ll H, ll W, vector<vector<ll>>& A) {
    vvpll B(M + 1);

    rep(y, H) rep(x, W) {
        if (A[y][x] != 0) B[A[y][x]].emplace_back(y, x);
    }

    ll ans = 0;
    for (auto& v : B) {
        if (sz(v) == 0) continue;
        if (sz(v) == 1) {
            ans++;
            continue;
        }

        Compress<ll> row, col;
        for (auto&& [y, x] : v) {
            row.push(y);
            col.push(x);
        }
        row.compress(), col.compress();

        ll h = row.size(), w = col.size();
        PSP<__int128_t> psp(h + w);
        auto var = psp.var;

        rep(y, h) psp.add_constraint(y, 1);
        rep(x, w) psp.add_constraint(x + h, 1);

        for (auto [y, x] : v) {
            y = row.zip(y);
            x = col.zip(x);
            psp.add_constraint({y, false}, {x + h, false}, 100000);
        }

        ans += (ll)psp.solve();
    }
    print(ans);
}


int main() {
    ll H, W;
    cin >> H >> W;
    vector<vector<ll>> A(H, vector<ll>(W));
    rep(j, H) {
        rep(i, W) {
            cin >> A[j][i];
        }
    }

    solve(H, W, A);
    return 0;
}